The activities of this task within the program were the studies of long-term systems. It was largely a research-and-development effort with activities concentrated on the accelerated development of promising technologies that are essential crash warning elements. The main focus has been on the systems and vehicle integration aspects of collision warning systems. Several demonstration vehicles, equipped with the rudimentary capabilities of a forward collision warning system, have been designed, developed, implemented, and successfully demonstrated. These vehicles demonstrated the viability of the baseline system architecture. Additionally, progress has been achieved in the development of a key strategic technology dealing with path estimation and target selection algorithm/software. The stated goals and objectives for this task are summarized as:
One of the goals of this task was to provide a viable flexible vehicular environment in which the activities of other program tasks could demonstrate, evaluate and assess the impact of their respective technologies within the framework of a complete collision warning system.
Three FCW demonstration vehicles were developed in order to support various collision warning development and evaluation activities of the ACAS Program. These three vehicles were: (a) 1994 Toyota Lexus LS400, (b) 1994 GM Cadillac Seville, and (c) 1995 GM GMC Suburban. These vehicles were modified to provide the basic functionality of a fully integrated FCW system. These FCW-equipped demonstration vehicles were developed by using an upward integrated design philosophy which included all aspects of the comprehensive FCW design architecture, from the FCW sensor, to the human factor designed set of driver vehicle interface (DVI) cues.
These FCW vehicles were primarily used to assess the "real-world" technical issues and challenges associated with integrated collision warning systems in support of the following program tasks:
Each vehicle had slightly different features and functionalities. However, the design approach followed a common architectural process, which is shown in Figure 3.17. The core of the system is the Collision Avoidance Processor (CAP) which takes the inputs from the sensor suite (active and vehicle sensors), processes the sensor information using the collision warning processing suites (i.e.: path determination, in-path target selection, threat assessment), and provide the appropriate driver-vehicle warning response. A data acquisition system is used to provide diagnostic features (i.e.: datalogging, etc.)
Figure 3.17: Collision Warning Vehicle Mechanization.
A functional description of each FCW demonstration vehicle is provided in detail in the Task 2.3 Summary Report. Each successive demonstration vehicle contained noticeably improved design features over the previous vehicle design. Lessons learned from on-road evaluations, and many of the comments provided by numerous subjects on each demonstration vehicle design, were examined and incorporated into the next demonstration vehicle design.
In order for a FCW system to provide a positive and beneficial influence towards the reduction of potential crashes, it is critical that the FCW system has the ability to correctly identify the vehicle/targets in the Host vehicle's path. The solution to this problem relies primarily on the FCW system's ability to estimate the relative inter-vehicular motion path (i.e.: range, relative speed, radius-of-curvature, etc.) between the Host vehicle and all of the appropriate targets (i.e.: roadside objects, vehicles, etc.), and on the system's ability to predict the mutual intersection of these motion paths. As one could imagine, the in-path target identification and selection problem is technically very complicated and challenging.
Figure 3.18 presents an illustration of the complexity of this problem. In this illustration, it shows a Host vehicle, equipped with a FCW system, which must correctly identify the in-lane target while navigating a random complex roadway segment in the random presence of complex driver/roadway events and rich target environment, while using realistic sensors. Some examples of the realistic driving environment characteristics that are presented to the Host vehicle are:
Figure 3.18: Collision Warning Path Prediction Problem.
In general, the critical scenarios which describe most inter-vehicle interactions can be grouped into three distinct categories, with the following geometric properties: (a) Straight Road Condition (i.e., Host and target vehicles are both traveling on a straight roadway); (b) Curved Road Condition (i.e., Host vehicle and target vehicles are both traveling on a curved roadway); and (c) Curved Entry/Exit Road Condition (i.e., roadway transition case in which the target vehicle is separately transitioning to another distinct curved or straight roadway). These roadway scenarios are illustrated in Figure 3.19.
Figure 3.19: Roadway Scenarios for In-Path Target Selection Process.
Consequently, the basic FCW path algorithm issue is to develop the appropriate decision logic to: (a) determine which geometric inter-vehicular scenario the Host vehicle is executing; (b) identify the complete set of target vehicles which are in the Host vehicle's path, based upon the identified inter-vehicle geometric scenario; and (c) select the in-path target which poses the highest threat to the Host vehicle.
The "conventional" design approach is briefly summarized as: (a) predict the Host vehicle path trajectory by using the passive in-vehicle sensors, and (b) identify the appropriate detected targets that are positioned within the path of the projected Host vehicle motion. In general, the "conventional" approach heavily utilizes the yaw rate information in order to estimate the roadway curvature.
Algorithm Development Activities
The development of this FCW path algorithm suite has been a lengthy process of continual refinements and enhancements in response to noted performance deficiencies observed during the extensive road testing procedures. In particular, the following significant improvements were made to the FCW system during the ACAS Program:
The algorithm development and evaluation efforts have focused on the three distinct roadway scenarios as previous depicted in Figure 3.19. In its present state, the conventional FCW path algorithm suite performance is excellent for simple scenarios, and good for complex scenarios. The performance for simple geometric scenarios (i.e., straight roadways with few targets, etc.) has shown a significant reduction in the number of false alarms on oncoming adjacent-lane targets and overhead bridges. Similarly, the performance for complex geometric scenarios has shown a substantial reduction in its rate of false alarms and missed detections.
Areas of further improvements which the FCW path algorithm suite still exhibits are: (a) missed detections at long range on transitions roadway scenarios and on rolling terrain (i.e.: bumps and hills); (b) false alarms on roadside object located at curved and transition roadways during both lane change and curve entry/exit maneuvers; and (c) false alarms at long range on adjacent lane cargo trucks. The root causes of these deficiencies can be attributed to the following issues:
1. Steering Behavior Variability: The observed random characteristics in the driving behavior still posses a significant issue. As a result of this driving behavior characteristics, it becomes difficult to rapidly assess the differences between a Host vehicle: (a) traversing a straight roadway segment, (b) initiating a lane change, or (c) initiating a "true" turning maneuver at a roadway transition (i.e.: curve entry/exit).
2. Target Glint Affects on Heavy Duty Vehicles. Target glint is indicative of MMW-based FLR sensors. This affect typically occurs when the aspect view of the target changes while on a curved road. For instance, on a straight roadway, the radar sensor "sees" a relatively constant radar cross section image of the target (i.e., constant aspect view of the vehicle trunk area). Consequently, the reported radar target centroid parameter attribute remains constant and stable. However, on any curved roadway, the radar sensor "sees" a dynamically changing aspect view of the target (i.e., vehicle trunk and side). When the target is a highly complex object (which is typical of most vehicles) or a large object, then the radar cross section can change dramatically between successive radar update cycle. Consequently, either the target can disappear, or the reported radar target centroid parameter attribute can become highly unstable (i.e., dramatic azimuth angle variations or shifts between successive radar update cycles). In particular, due to the glint phenomena, large adjacent-lane vehicles (i.e.: trucks, etc.) tend to provide a higher false alarm rate, in comparison to passenger-style vehicles. Moreover, the false alarms on the large vehicles tend to occur more frequently at long range, on transition roadways (i.e., curve/entry exit roadways). In general, it is possible for the centroid of a typical heavy-duty vehicle to vary by an amount greater than the real vehicle's width. This variation is due to the glint characteristics of the vehicle's numerous sharp reflective corners. Figure 3.20 illustrates centroid variations for three different types of vehicles (i.e., passenger, dump truck and rollback wrecker). The Host vehicle is passing by the adjacent-lane vehicle, while traveling on a straight roadway. The Host and adjacent-lane vehicles are traveling at 84 MPH and 65 MPH, respectively. The inter-vehicle spacing varied from 150 meters to about 20 meters. The standard deviations in the centroid position of the passenger car, dump truck and rollback wrecker were 0.42 meter, 0.52 meter and 0.76 meter, respectively. Moreover, throughout the test run, the maximum variations of the target centroids, for the passenger car, dump truck and rollback wrecker were 1.75 meter, 2.5 meter and 3.2 meter, respectively. This maximum variation roughly corresponds to the respective width of each of the vehicles.
Figure 3.20: Centroid Variations for Three Types of Vehicles as a Result of Target Glint.
3. Inadequate Filtering of Target Attributes and Host Vehicle Yaw Rate. The filtering of the Host and target vehicle attributes is used to eliminate variability in driver steering behavior and yaw rate signal noise. However, if the filtering is too harsh, then the FCW system exhibits too much lag and respond slowly to normal dynamic vehicular maneuvers (i.e., lane changing, changing roadway curvature, and straight/curve roadway transitions, etc.). During the ACAS Program, the filtering scheme for the Host and target Vehicle's attributes were enhanced to reduce the effects of driver steering variability, and to increase the system's responsiveness to close-range cut-in targets and Host lane change maneuvers. Figure 3.21 presents a comparison of the two types of filter design implementation for the Host vehicle yaw rate parameter attribute. The "old" filter design did not provide adequate filtering characteristics; in fact it provided very little filtering action in comparison to the "new" (i.e., current implementation) filter design. However, while the new filtering scheme significantly improved the FCW system's performance for many of the critical in-vehicle scenarios, the current approach does not provide enough responsiveness to target vehicles performing very rapid lane change maneuver (i.e., cutting-in and cutting-out of the lane of the Host vehicle at a high rate of speed), and to target vehicles performing curve entry/exit maneuvers at a high rate of speed. One possible way to improve the system's responsiveness may be to filter the target attributes and Host vehicle yaw rate using some adaptive scheme. In such a suggested scheme, sets of filter coefficients could be selected based on such system attributes as the yaw rate angular acceleration, and the speed of the target and Host vehicles.
Figure 3.21: Yaw Rate Filter Design Characteristics Comparison.
4. Collective FCW System Sensor Suite Limitations: The conventional design approach relies on the use of a single active detection sensor (e.g., radar) and a single passive in-vehicle sensor (e.g., yaw rate, steering, and speed) in order to identify the in-path target. However, these collective set of sensors do not readily provide lateral placement in the lane or discern lane and road boundaries. In the absence of lane boundary or roadway curvature information in the area ahead of the Host vehicle, it will be very difficult to reliably anticipate/predict:
- Changes in the roadway curvature (i.e.: curve entry/exit transitions),
- Differentiate between target lane-change maneuvers and curve-entry/exit maneuvers.
- Differentiate between Host vehicle lane-change maneuvers and curve-entry/exit maneuvers.
- Determine that roadside objects (i.e.: signs, poles, parked vehicle, etc.) which are located the curve-entry/exit point do not lie along the Host vehicle path.
- Identify if either the Host or target vehicles are hugging the edge of their respective lanes.
Consequently, in face of these challenges, it will be difficult for the currently implemented conventional FCW path algorithm design approach to correctly select in-path targets (or reject adjacent-lane target) during roadway transition scenarios, or reject adjacent-lane line-hugging vehicles as in-path targets.
At this time, it is believed overall system performance can be further enhanced by augmenting the current conventional FCW path algorithm design approach by incorporating a model-based scene tracking estimation technique (see discussion in the next sub-section), and/or incorporating a hybrid sensor architecture (i.e.: multiple active detection sensors), as suggested by the CW system architecture presented in Figure 1.1. For example, a vision system, which has the ability to provide an estimate of the road curvature and boundaries ahead of the Host, could assist in enhancing the system performance.
Presently, the performance of the conventional FCW path algorithm design approach is overall very good. It provides excellent performance in the presence of simple roadway/driver scenarios, and good performance for complex roadway/driver scenarios. This is a significant improvement when compared to the level of performance at the program inception. The performance for simple geometric scenarios (i.e., straight roadways with few targets, etc.) has shown a significant reduction in the number of false alarms on oncoming adjacent-lane targets and overhead bridges. Similarly, the performance for complex geometric scenarios has shown a substantial reduction in its rate of false alarms and missed detections. In addition, the duration of over 95% of the missed detections, and over 90% false alarms, has been reduced to less than 0.5 seconds.
In this sub-section, several figures are presented as a means to provide a qualitative comparison of performance improvements of the FCW path algorithm suite. The comparisons will correspond to three distinct time periods during the ACAS Program: (1) "new" algorithms at the completion of the program (December 1997), (2) "old" baseline algorithms developed at the fifth quarter (February 1996), and (3) "intermediate" algorithms developed at the seventh quarter (October 1996). These figures depict actual real-time FCW path algorithm performance in a "real-world" driving environment for a variety of roadway events. They will demonstrate the following performance improvements: (i) reduction in the duration and frequency of false alarms (i.e.: out-of-lane targets incorrectly identified as in-path targets) due to adjacent-lane vehicles or roadside objects, and a reduction in the frequency of missed detections (i.e.: in-lane targets not identified as in-path targets) due to driver lane hunting/wandering; (ii) significant reduction in the number of false alarms triggered by opposing traffic on a two-way surface streets; (iii) development of a new and effective bridge discrimination capability (i.e.: ability to differentiate between vehicular objects and overhead freeway bridges and stopped surface objects; (iv) reduction of the average response time for target vehicle cut-in and Host vehicle lane change maneuver; and (v) reduction in the duration of 90% of the FCW systems false alarms and missed detections to an average duration of between 0.1 and 0.2 seconds.
Simple Geometric Scenarios
Figure 3.22 shows the progression in the system performance on a simple geometric roadway scenario, in which two moving vehicles (i.e.: Host vehicle and target vehicle) are traveling on a straight roadway segment. The target and Host vehicle are both traveling in the same lane with a constant velocity of approximately 50-MPH. The separation between the vehicles is about 55 meters. Multiple isolated sign poles are also located along the Host/Target vehicle's lane edge (i.e., roadside).
Figure 3.22: Performance Comparison (Lane Hunting/Wander Scenario).
During the course of the test drive, the Host vehicle driver simulated various lane-hunting maneuvers, which resulted in yaw rate fluctuations of approximately 0.5 deg/sec. This figure shows that the "old" (Feb 1996) FCW path algorithm suite incorrectly selected many of the sign poles as in-path targets (i.e.: false alarms), and also failed to identify the primary in-path target vehicle on several occasions (i.e.: missed detections). As a comparison, the in-path target selection performance was significantly improved with the use of the "new" (Dec 1997) FCW path algorithm suite. The improved algorithm suite eliminated all the false alarms and all the missed detections that were observed using the "old" baseline algorithm suite.
Figure 3.23 compares the system performance for a simple geometric roadway scenario, in which two moving vehicles (i.e.: Host vehicle and target vehicle) are traveling in the same lane on a straight surface roadway with two-way traffic. There were also several on-coming vehicles approaching in the opposite direction. All of the moving vehicles are traveling between 40 MPH and 50 MPH. The separation between the Host and the lead vehicle varies from 30 to 120 meters. This figure shows that the "old" baseline FCW path algorithm suite exhibited many false alarms by selecting the oncoming adjacent-lane vehicles. On the other hand, the "intermediate" FCW path algorithm suite exhibited only a single false alarm on the oncoming adjacent-lane vehicle, while the "new" improved FCW path algorithm suite eliminated all occurrences of the false alarms experienced by the "old" baseline approach.
Figure 3.23: Performance Comparison (Oncoming Adjacent-Lane Vehicles Scenario).
Figure 3.24 compares the system performance for a simple geometric roadway scenario, in which the Host vehicle is traveling on a straight section of a freeway with an overhead bridge. The Host vehicle is approaching an overhead bridge at 65 MPH. The bottom of the overhead bridge is located approximately 4.5 meter above the roadway surface. The radar sensor detects the overhead bridge as a stopped object at a range from 125 meters to 75 meters (i.e.: with in the elevation field-of-view of the radar sensor). This figure shows that both the "old" baseline and "intermediate" FCW path algorithm suite incorrectly selected the overhead bridge as an in-path surface target for a duration of about 1.3 second and 0.6 second, respectively. However, the "new" improved path algorithms successfully recognized the overhead bridge as a non-vehicular object, and rejected it as an in-path target.
Figure 3.24: Performance Comparison (Overhead Bridge Scenario).
Complex Geometric Scenarios
Figure 3.25 compares the system performance for a complex geometric roadway scenario, in which three moving vehicles (i.e.: Host vehicle and two target vehicles) are traveling on a curved roadway with a counter-clockwise curvature of 2000 meters. One of the target vehicles is in the same lane as with the Host vehicle, while the other target vehicle is in the right adjacent-lane. The Host vehicle is traveling at a constant speed of 60 MPH, while the two target vehicles are traveling next to each other at a constant speed of 55 MPH (i.e.: closing speed of 5 MPH). The inter-vehicle separation distance, between the Host vehicle and both target vehicles, is varies from between 100 to 60 meters.
Figure 3.25: Performance Comparison (In-Lane & Adjacent-Lane Vehicles Scenario).
This figure demonstrates the "old" FCW path algorithm suite frequently failed to detect the in-lane target vehicle as the primary in-path vehicle. Moreover, on one occasion, the right adjacent-lane target vehicle was incorrectly identified as the primary in-path target. In comparison, the "intermediate" FCW path algorithm suite did not experience any false alarms due to the presence of the right adjacent-lane target vehicle. Moreover, the number of missed detections was reduced dramatically. The in-path target selection performance was significantly improved with the use of the "new" refined FCW path algorithm suite. The improved algorithms eliminated all of the false alarms and missed detections for this scenario. As a result, the correct in-path vehicle was selected throughout the entire test run.
Figure 3.26 compares the system performance for a complex geometric roadway scenario, in which two moving vehicles (i.e.: Host vehicle and target vehicle) are traveling on a multiple-lane curved roadway with a straight road transition. The curved roadway has a radius of curvature of 1800 meters. Initially, both vehicles are traveling at about 65 MPH. The target vehicle is traveling on the left adjacent lane relative to the Host vehicle, and is approximately 40 meters ahead of the Host vehicle. The target vehicle accelerates to 70 MPH, and then cuts into the Host vehicle lane. This figure shows that the responsiveness of the FCW system to this cut-in maneuver varied significantly between the various versions of the FCW path algorithm suite. The figure demonstrates that the "old" FCW path algorithm suite exhibited a delay of 0.4 seconds and 1.2 seconds in response to this target vehicle cut-in maneuver, as compared to the "intermediate" and "new" algorithm suite. This quick response is very significant because 60% of the rear-end crashes could potentially be avoided if the driver has an extra 0.5 seconds in reacting to the situation.
Figure 3.26: Performance Comparison (Target Vehicle Cut-in Scenario).
Figure 3.27 compares the system performance for a complex geometric roadway scenario, in which two moving vehicles (i.e.: Host Vehicle and in-lane vehicle) are traveling on an S-curved roadway with a straight road transition. The first part of the S-curve has a radius of curvature of 1500 meters, while the second part of the S-curve has a radius of curvature of 1100 meters. Both the Host and target vehicles are traveling at 45 MPH, with the inter-vehicle separation distance of 50 to 70 meters. This figure demonstrates the "old" FCW path algorithm suite frequently false alarmed on the poles along the second part of the S-curve (i.e.: incorrectly identified the poles as in-path targets). Moreover, it also failed to identify the in-lane target vehicle as an in-path target on several occasions. On the contrary, the "intermediate" algorithm suite did not experience any false alarms due to the presence of the sign poles at the S-curve transition. Moreover, the number of missed detections was reduced to one with duration of only 0.2 seconds. The in-path target selection performance was significantly improved with the use of the "new- refined " FCW path algorithm suite. The improved algorithms eliminated all false alarms and missed detections. As a result, the in-lane vehicle was correctly selected throughout the entire test run.
Figure 3.27: Performance Comparison (In-Lane Vehicle on S-Curve Scenario).
Observations & Comments
Significant progress has been achieved in developing the "conventional" path determination/estimation algorithm suite in addressing the collision warning target selection function. Areas of future activities and investigations are: (a) improving the FCW path algorithm suite performance in response to complex geometric scenario events, and (b) validating the consistency and robustness of the FCW path algorithm suite in response to various different performing yaw rate sensors and Host vehicle platforms.
As discussed in the previous section, the currently implemented conventional FCW path algorithm suite design approach does not provide the capability to reliably anticipate/predict changes in roadway curvatures ahead of the Host vehicle (i.e.: roadway transition scenarios), due to its inability to readily discern lane and road boundaries. Consequently, it is difficult to correctly select the actual in-path target (or reject adjacent-lane target) during roadway transition scenarios. Figure 3.28 illustrates a typical scenario in which the current conventional FCW path algorithm suite implementation might have difficulty in identifying the correct in-path target. In this scenario, the Host vehicle is operating on a straight roadway segment, while the in-path target has entered onto a sharp curved roadway segment. In general, the current conventional algorithm implementation heavily utilizes the yaw rate in order to estimate the instantaneous roadway curvature at the Host vehicle. As such, while the Host vehicle is on the straight roadway segment, the current algorithm implementation would accurately estimate the roadway curvature to be a straight roadway, but would not provide any indication of a dynamically changing roadway curvature ahead of the Host vehicle. In general, the current algorithm implementation does not fully utilize the active detection sensor (i.e.: radar, laser, vision, etc.) generated data (i.e., range, range rate, angle, etc.) to assist in estimating the roadway curvature. Consequently, for this scenario, the current conventional FCW path algorithm suite might incorrectly select the roadside light pole or sign as the in-path target, rather than the proper actual vehicle.
Figure 3.28: Typical Roadway Scenario.
In order to address this issue in the near term, another FCW path algorithm design approach was investigated to fully utilize all the information from the active sensor. This has led to the development of a vehicle/roadway model-based estimation technique, which provides a dynamic estimation/prediction of the road curvature ahead of the Host vehicle by tracking the position and trajectory of all of the detected objects within the active sensor field-of-view. By using all of the relevant data collected by the forward-looking active sensor (i.e., radar or laser), and other in-vehicle passive sensors (e.g., speed, yaw rate, etc.), it is possible to dynamically reconstruct the scene ahead of the Host vehicle in order to significantly improve the in-path target selection process. This scene tracking technique makes much greater use of the available scene information and sensor data than does the conventional FCW path algorithm suite.
In the future, this issue could also be addressed by the integration of a vision sensor which provides an improved estimate of the roadway curvature and boundaries ahead of the Host vehicle. In addition to the potential for improved performance, the model-based scene tracking approach provides a natural setting for the fusion of data from multiple forward- looking sensors (i.e.: radar and vision together).
Simulation Environment
The model-based scene-tracking design approach was initially developed in a simulation environment, rather than as a direct real-time vehicle implementation. This simulation environment provides an ideal setting from which controlled, repeatable evaluation of system performance, and sensitivity to alternative configurations, scenarios and error sources can be performed without having to deal with real-time vehicle implementation considerations. Moreover, this simulation can be used to investigate the use of different combinations of on-board sensor data (e.g., yaw rate, steering angle, wheel speeds, lateral acceleration, forward-looking sensor data, etc.) in order to understand their influence on possible FCW path performance improvements.
The Simulation Environment consists of several modules which provide the following capabilities: (a) generate an arbitrarily prescribed roadway, generate trajectories for an arbitrary number of targets; (b) steer the Host vehicle with realistic driving dynamics along an arbitrarily prescribed trajectory; (c) model the collection and corruption of sensor data (including the forward-looking radar); and (d) pass the sensed data to a path algorithm under study, and perform data logging. The simulation is implemented using the software package Matlab™, operating on a PC platform.
The simulation assumes that the Host and multiple target vehicles are traveling down a multi-lane road of a prescribed shape. The simulation allows the user to independently specify the desired road scenario, on-board sensor configuration, and kinematic behavior of the vehicles. Moreover, the simulation includes both a forward-looking radar model, and models of other in-vehicle sensors (speed, yaw rate, steer angle, etc.). The radar model simulates a variety of realistic major radar parameters, such as: number of detected targets, detection range, field-of-view, resolution and quantization of the radar's target features (range, range rate, azimuth angle), and various other error sources. Furthermore, the characteristics of the radar model is representative of the Delco Electronics/HEM radar sensor that has been used in the ACAS Program.
Model-Based Scene Tracking Algorithm Architecture
The structure of the model-based scene-tracking algorithm is shown in Figure 3.29. It analyzes the features of the Host vehicle (i.e., speed, yaw rate, steering angle) and "n" detected targets (i.e., range, range rate, and azimuth angle). The algorithm has four main components:
Figure 3.29: Model-Based Scene Tracking Algorithm Architecture.
The Host's path angle is the angle between the Host's longitudinal axis and the tangent to the local lane center, as depicted in Figure 3.30.. It provides a measure of the extent to which the Host vehicle isn't pointing parallel to the nearby road. In the vehicle/roadway model, the path angle indicates the direction in which the estimated nearby road segment should be projected, relative to the Host (i.e., not necessarily straight-ahead), for the target Lane Position Estimation process.
Figure 3.30: Path Angle Definition.
Algorithm Performance
The scene tracking algorithm was fine-tuned to maximize its performance with regard to anticipated non-idealities of "real-world" driving/roadway such as: (a) weaving of Host and target vehicles in their respective lanes; (b) lane changes by Host and target vehicles; (c) variable number of targets; (d) varying speeds of Host and targets; (e) complex roadway scenarios (i.e.: straight, curve, transition to curve); and (f) sensor characteristics (i.e., sensor misalignment; target dropouts, or momentary loss of target data due to limited field-of-view, glint and scintillation).
A "real-world" driving/roadway event which can robustly be handled by the scene-tracking algorithm is the in-lane weaving of both the Host and target vehicles. Figure 3.31 presents a scene tracking algorithm comparison for a complex geometric roadway scenario, in which four moving vehicles (Host Vehicle and 3 target vehicles) are traveling on a roadway which transitions from a straight to a curved road, with a constant 800 meter radius-of-curvature. In this event, all the vehicles are traveling at 20 m/s. A target vehicle is present in the left adjacent, right adjacent and same lane of the Host vehicle. The target orientations are: (a) left adjacent lane vehicle (Target 3) is located 70 meter ahead of the Host vehicle; (b) same lane vehicle (Target 2) is located 90 meter ahead of the Host vehicle; and (c) right adjacent lane vehicle (Target 1) is located 50 meter ahead of the Host vehicle. This figures compares the lane position estimation errors for each target, for the following two driving cases: (i) Neither Host nor targets are weaving; and (ii) Both Host and all targets are weaving (i.e.: weaving characteristics: sinusoidal, harmonically unrelated periods, and 1m peak-to-peak amplitude). As it can be seen, the scene tracking system provides very acceptable lane position estimation performance in the presence of weaving. As would be expected, the worst performance was for the most distant target (target 2 at 90m). In fact, for the near-range targets (targets 1 & 3), very little performance differences are observed between the "weaving" and "non-weaving" scenarios.
Figure 3.31: Weaving Scenario (Curve Transition vs. No Transition).
One interesting and intuitively reasonable finding of the scene tracking development effort is that a performance tradeoff is required between: (a) detecting lane changes, (b) maintaining proper target lane classification in curve entry & exit scenarios, and (c) maximizing insensitivity to in-lane weaving of targets. This is particularly true when only one target is visible in the scene, since there are no other targets available to corroborate information related to the shape of the upcoming road segment.
One aspect of this tradeoff can be seen in Figure 3.32. This figure depicts two plot windows which shows a comparison between the estimated and actual target lane positions for a two different single target roadway scenarios. The characteristics of the target/roadway events are: (a) Single target traveling on straight roadway segment and changes lanes [See Top Plot]; and (b) Single target traveling on a straight roadway segment which abruptly changes in a 800 meter radius-of-curvature turn and remains in the same lane (i.e. no lane changes) [See Bottom Plot]. In the lower plot window, it can be seen that the target's lane position is accurately tracked as it enters the sudden turn. This indicates that the scene-tracking algorithm is rapidly changing its estimate of the road shape, while allowing the target to remain centered in the lane. In upper plot window, it shows that as a consequence of this rapid road shape adaptation, it takes a few seconds to realize that the target changed lanes on the straight road.
Figure 3.32: Comparison of Lane Change vs. Turn Entry Maneuver.
Figure 3.33 clearly illustrates what is happening in the lane change scenario described in Figure 3.32. Five overhead view snapshots are shown covering the interval of time (i.e.: time from 40 to 44 seconds) which completely encompasses the lane change. Each overhead snapshot shows the front end of the Host vehicle at the bottom of the picture, and the target near the center. The estimated target position, as calculated by the Target Tracking Filter, is shown as a small circle inside the box representing the target. The Host's estimate of the upcoming road shape is shown as a dashed line. Prior to any lateral motion of the target (i.e.: t = 39 sec), the Host's estimate of the road shape is straight and directly through the center of the target. In the subsequent two snapshots (i.e.: t = 41 and 43 sec), the target is seen to be changing lanes, and the Host's estimate of the road is bent to follow the target into the next lane. At t = 45, the target is now going straight again, and the Host's road shape is again bent to account for that. There is now a nonzero path angle estimate evident (i.e.: the estimated road doesn't come straight out of the front of the Host). After seeing the target go straight for several seconds (i.e.: t = 49 sec), the algorithm now realizes that the road is actually straight and that the target has changed lanes.
As might be expected, a lane change by one target can be more easily recognized if there are more than one target in view. This is due to the greater amount of road curvature information that is available from two or more targets.
Figure 3.33: Evolution of Upcoming Road Segment Estimate.
Figure 3.34 reveals another interesting problem which merits further attention, which involves the target lane change event, for a simple geometric roadway scenario, in which three moving vehicles (i.e.: Host Vehicle and 2 target vehicles) are traveling on a straight roadway segment. In this event, all the vehicles are traveling at 20 m/s. The upper plot window depicts a two-target straight road scenario in which the near target makes a lane change. Similarly, the lower window depicts a two-target straight road scenario in which the distant target makes a lane change. The lower plot window shows that when the distant target makes a lane change, the lane change is recognized with less delay than in the single target case, and the target which does not change lanes is tracked very well. On the other hand, in the upper plot window, the near target lane change causes the upcoming road estimate to be projected into the adjacent lane, which makes it briefly appear that the distant target is making a lane change in the other direction. This apparent range-dependent influence on road curvature estimation requires further study.
Figure 3.34: Straight Road Scenario: Distant vs. Near Target Lane Change.
Comparison between Scene Tracking and Conventional Yaw Rate Based Path Estimation
As discussed previously, one of the benefits of using a simulation environment is that controlled performance comparisons of different algorithms can be conducted. In general, the scene tracking approach has better tracking capability in the presence of targets performing sudden curve entry maneuvers, and can better deal with steering maneuvers made by the Host (weaving and lane changes). However, the scene tracker also has more difficulty in handling the scenario of only weaving targets.
Figure 3.35 illustrates the errors in estimating the target lane positions, as a comparison between the "scene tracker" and the "conventional" path algorithms. This scenario involves a single target and the Host vehicle, where the roadway is straight and then abruptly changes into a 500m constant radius-of-curvature turn. The target is 90 meters ahead of Host vehicle and is located in the same lane. In this figure, the conventional algorithm, sensing that the road is straight (since the Host vehicle is still traversing the straight road segment), initially believes that the target has changed lanes (since the target has entered the curve). Only after the Host vehicle has also entered the curve, and is well into the turn, then the "conventional" algorithm will again correctly classifies the target as in-lane. On the other hand, the scene-tracking algorithm is only minimally bothered by the abrupt transition in the roadway.
Figure 3.35: Comparison Approaches (Host Remaining In-Lane without Weaving).
Observations & Comments
The model-based scene tracking approach offers the potential for improved path estimation performance in a number of areas in which the conventional approach has deficiencies. Moreover, it also provides a basis for the fusion of data from a variety of sensors. However, several topics related to scene tracking require further study. For example, an approach to integrate and utilize both scene tracking and conventional path estimation techniques to provide added robustness and redundancy to the system needs to be developed. A tradeoff analysis between recognition of target curve entry, timely detection of lane changes, and immunity to weaving of targets needs to be carried out. The relative influence of targets at different ranges upon road curvature and path angle estimates needs to be investigated. The extension of the scene tracker to handle variable speeds of Host and target vehicles, the steady state implementation of all the filters, and the investigation of the potential for using the scene tracker attributes to estimate yaw rate sensor bias and forward-looking sensor misalignment should also be tackled.