3.2 Development of Near-Term Systems (Task 2.1)

Various studies of crash scenarios have indicated that over one-quarter of the police reported crashes are rear crashes with other vehicles or objects. In addition, a large percentage of these crashes are due to an inattentive or distracted driver. If an appropriate warning could be provided to an inattentive or distracted driver, a number of these crashes may be avoided. This warning should be appropriate to the driver. That is, it should identify a true potential crash situation and should avoid nuisance alerts. Nuisance alerts could be generated too early (i.e. before the driver would consider the situation a potential crash) or due to objects which are not in the vehicle's path. Nuisance alerts may cause the driver to become insensitive to the warning and thereby miss an appropriate situation. To provide a warning of a potential rear end crash, the warning system requires a forward-looking sensor. This sensor and system should be cost-effective for the automotive market.

This task identified the requirements for a rear-end crash warning system and sensor based on the studies performed in Task 1 - Overall Program Requirements and Performance Validation. The system elements that were identified included sensor, processor, and warning device. The primary effort was devoted to identifying and evaluating a forward-looking sensor. Radar and optical sensors were researched to determine the viability of each technology for this application. The radar technology was selected because of its ability to identify objects in the forward path in rain, snow, and fog conditions. Sensor parameters required for the warning application were identified and a sensor satisfying these requirements was purchased, installed on a vehicle, and evaluated. Threat assessment algorithms, which determined the potential of a rear end crash with an object in the vehicle's forward path, were identified and were installed in the crash warning processor. Roadway evaluations of the sensor and warning algorithms were performed.

The goals and objectives of this task were:

• To derive the requirements for the rear end crash warning system and define sensor specifications
• To obtain a developmental radar sensor from a component supplier
• To integrate the sensor into a vehicle and evaluate the rear end crash warning concept with in-vehicle tests
• The expectations of this task were to verify that the purchased forward-looking radar sensor met the sensor requirements which were established from crash scenarios and normal driving situations and to demonstrate threat assessment warnings due to objects identified by the radar sensor. A number of test scenarios were identified that would verify the capability and its ability to satisfy the defined requirements. Any deficiencies would be analyzed and determined whether the requirements should be adapted or the sensor design should be modified.

3.2.1 Sensor Parameter Requirements and Rationale

Based on the Forward Collision Warning (FCW) system requirements defined in Task 1, key forward-looking sensor parameters were identified as important characteristics in future collision warning systems. Initial requirement values for each parameter have been determined based on preliminary analyses and are shown in Table 3.2.

These parameters and values were used as the basis for selecting a potential developmental radar sensor for further evaluation on vehicles and in test situations. These requirement values, however, should be considered as initial engineering estimates for forward-looking sensor requirements pending further analysis, simulation, and vehicle testing. The rationale for selecting these values is:

• Azimuth Field of View (FOV): Sensor is required to detect a stopped object in the same lane at 150 meter range while on a 500 meter radius curve. This also results in fully illuminating the adjacent lane at a range of 100 meters.
• Elevation Field of View (FOV): Sensor must be able to keep track of objects which are within range and azimuth FOV and account for road tilt (5% grade), road variation, sensor misalignment, and vehicle pitch. Further analysis and field testing are required to determine a suitable value for the elevation FOV parameter.
• Operating Range: Sensor is required to detect/track stopped objects at a range that provides time for driver reaction. Minimum range of one car length (5 meters). Maximum range of 200 meters is based on a stopped object:
  a. at 65 mph (29 m/s) closing rate requires 150 m stopping distance (0.3 g deceleration),
  b. with total latency of 1 to 1.5 seconds results in additional 30 to 45 meters.
• Range Rate: Needs to be large to avoid aliasing or dropping target tracks. Maximum closing rate of no more than 100 m/s (220 mph). Maximum opening rate of 50 m/s (110 mph).
• Azimuth Resolution: Needs to accurately determine if an object is in the current path even if multiple vehicles are at same range and speed. Requires accurate position/range/velocity measurement within a range of 100 meters. Requires ability to resolve objects in adjacent lanes at 100 meters.
• Range Resolution: Requires multiple resolution cells on the object without overburdening the processor. Resolution equal to one-half a motorcycle length is the criterion (< 1.0 meter)
• Range Rate Accuracy: Needs to determine an object's velocity to within 0.25 m/s to maintain a desirable track file.
• Data Update Rate: Dynamic environment dictates that the sensor data be updated in a timely fashion. If two vehicles are approaching at 130 mph (58 m/s) then the range changes by one vehicle length (~5 meters) in 90-100 ms (~10 Hz update rate).
• Maximum Sidelobes Level (1-way): Nominal vehicle signatures can vary by more than 40 dB. High sidelobes will result in small objects being masked or large objects creating ghost (false) objects. Low sidelobes will also help reduce the RF interference.

3.2.2 Prototype Radar Sensor Evaluation and Results

Base on these sensor parameter requirements, a mechanically, scanning radar sensor was selected from a number of potential forward-looking radar sensor suppliers. The field of candidates was narrowed down to mechanically scanned sensors because these sensors satisfied the azimuth resolution and the azimuth field of view requirements. With an azimuth resolution of less than 1.5^o, the sensor theoretically should be able to resolve two adjacent vehicles one half of a lane width apart at a range of 100 meters. With an azimuth field of view of 18^o (9^o on either side of the forward line of sight), the sensor should be able to detect a vehicle in an adjacent lane at 100 meters and detect a stopped object in the same lane at 150 meters when traversing a curved path with a 500 meter radius of curvature.

The selection of the prototype sensor supplier was based on the specified performance of the candidate sensor and the development environment available for data acquisition and algorithm development. The selected sensor met all but one parameter requirement. Its development environment provides data acquisition and algorithm evaluation support which enables GM to develop threat assessment algorithms and test the algorithms in traffic situations. The sensor and development system have multiple stationary and moving object tracking capability, access to data at various stages in the processing chain, and extensive road test experience in an intelligent cruise control application.

The sensor was initially tested and evaluated in the laboratory environment to verify basic operational characteristics. After verification of its functionality, the sensor was installed in a GMC Suburban with a collision warning controller, threat assessment algorithms, and data acquisition equipment. Controlled static and dynamic tests were performed with the radar sensor on the vehicle.

One of the goals of this task is to implement the Forward Collision Warning (FCW) system on a vehicle to obtain field data to gain real world understanding of the problem. The overall FCW system consists of an instrumented vehicle with sensors, computing elements, a data acquisition system, a driver information system, an engineering terminal, an on-vehicle development environment, a video camera, and a video recorder. The most critical element in the FCW system is the forward-looking sensor, and radar was selected as the primary sensor. This type of sensor has the ability to sense objects in a limited volume in the front of the vehicle. There are several secondary vehicle dynamics sensors to supplement this primary sensor. The sensor suite in its current form is redundant, however it enables various alternatives to be evaluated for a robust FCW system implementation.

Computing elements consist of processing sub-systems that are dedicated to the forward-looking sensor and are usually provided by the sensor supplier. They are either special purpose systems designed and built for sensor signal processing or personal computer based systems, with optional special purpose add-in boards. They are also supplemented by a special purpose FCW computer to extend and enhance the functionality of the sensor.

A data acquisition system is integrated into the Forward Collision Warning computer with a laptop personal computer for permanent data storage. This system simultaneously collects radar data, vehicle dynamics data and video for off-line analysis. A driver information system is a rudimentary device that gives feedback to the driver/experimenter on the status of current driving situation. An engineering terminal and an on-vehicle development system are implemented on a laptop personal computer. This enables the operator to control the system, observe the results, and make necessary improvements.

Two vehicles have been instrumented with radar sensors to be used as testbeds for evaluation. Early in the program a commercial off-the-shelf system (COTS) is installed in a vehicle to gain experience. Later, a carefully selected Prototype sensor was acquired and installed in a vehicle for evaluation. The goals of the prototype system are to select a radar sensor, which exceeds the specifications of a production system sensor, to conduct in-depth analysis of the sensor, and to come up with sensor specifications for production unit. The production sensor specifications are expected to be less stringent than the prototype sensor.

A test vehicle has been instrumented to evaluate the prototype sensor as well as FCW system on test tracks and on real-world traffic. This is the testbed similar to the COTS system implemented on a Chevrolet Suburban. The block diagram of the Prototype Radar system architecture is shown in Figure 3.11.

The Prototype sensor has enhanced specifications compared to the COTS sensor. The most apparent characteristic of the prototype sensor, compared to the COTS sensor, is the size, mainly due to the scanning technology. Mechanical scanning is used with single antenna compared to four distinct antennas. In addition, the same antenna is used for transmit and receive compared to separate antennas in the COTS sensor. The beam scan rate is lower, 10 Hz, but with a wider field of view and a narrower beam.

Tests were conducted to evaluate the sensor parameters. No specific test has been performed to evaluate the overall FCW system. All of the tests were performed on restricted test areas or test tracks. The results reported here are for the prototype sensor only. The tests are classified in three groups; static, semi-static, and dynamic. The purpose of the static tests is to measure certain radar sensor parameters that are hard to measure and verify when the environment is changing and the radar equipped vehicle is moving. Important radar parameters such as range accuracy and resolution, and azimuth accuracy and resolution were evaluated using the static tests. Semi-static tests are performed such that target(s) is stationary and the radar-equipped vehicle is moving. These are used for latency, range rate accuracy, and angular resolution in a real world setting. In the dynamic tests, both the target(s) and the radar equipped vehicle are moving. It is the hardest type of tests to control various parameters, especially when there are multiple target vehicles. Also, it is difficult to repeat exactly the same scenario consistently.

The tests were performed to determine the compliance of the radar with the specifications in a real world environment. The results of this testing are summarized below:

1. Range accuracy test results were satisfactory and this parameter is within the specifications. However, range resolution results showed that this parameter was out of specifications.
2. Field of view and accuracy tests showed that the radar meets the specifications. Again, angular resolution measurements were out of specifications. Range and field of view measurements were performed using corner reflectors initially, then with real vehicles.
3. Range rate accuracy test proved that this parameter was within the specifications. Vehicle speed was used as a reference, which could have some error also.
4. Range rate latency measurement is used to determine the overall latency of radar output computation. However, it also includes the communication delay associated with data acquisition. The measured value is as expected.
5. Two stationary vehicles in adjacent lanes demonstrate the real world performance of the radar's angular resolution performance. The results indicate that these two vehicles are resolved at 40 meters that is well below the expected performance based on the specifications.
6. When two vehicles are moving at the same speed in adjacent lanes, the radar is capable of resolving them at 70 meters when closing in and 80 meters when opening. This is much better than in the stationary case but still does not meet the specifications.
7. In the case where these two vehicles are moving at different speeds in adjacent lanes, they are resolved at 100 meters which meets the specifications. This result indicates that the radar requires range and range rate parameters of two adjacent targets to be different to be able to resolve them.
8. The target cut-in test verified that the radar meets the specifications for the field of view parameter under dynamic conditions.
9. When two vehicles are moving at the same speed in two outside lanes, the radar is capable of resolving and tracking them at 170 meters. Also, when the radar vehicle approaches them and passes in between them, they are tracked up to angles of +8 and -9 degrees. These two results are well within the specifications of the radar.

In general, the radar was within the specifications during some of the tests. Those cases that the radar met the specifications were observed when the targets were moving and there was at least one parameter varying among multiple targets. In those cases where targets were stationary or all the parameters were the same, except the measured one, the radar failed to meet the specifications. In addition, the imbedded sensor processor first processes the parameters that were measured which may involve some filtering. Thus, they do not represent the raw data measured by the radar. This processing is mainly for adaptive cruise control and/or forward collision warning type of applications.

3.2.3 Simulation Approach and Results

The purpose of the Forward Collision Warning System (FCWS) simulation is to develop an engineering tool to assist in the definition of the system requirements and refinement of the functional radar sensor and to evaluate technical and functional specifications of the radar sensor. The simulation incorporates roadway geometry, traffic scenario, and radar sensor model for analysis and display. The simulation also provides a tool for analyzing different forward collision warning algorithms under normal and crash scenarios. It allows evaluation of threat assessment algorithms in repeatable simulated scenarios and performs sensitivity analyses of system and sensor parameters to substantiate specifications.

The system simulation consisting of an interactive driving model, a radar sensor model including a simplified object tracking model, an integrated threat assessment module, and the user display module. The simulation, an engineering tool, generates positional information of the host and other moving and stationary objects in the scene, host vehicle dynamics, measured radar parameters, tracked object histories, and a time to collision parameter for various traffic scenarios. The concept of the simulation is shown in Figure 3.12.

Different host vehicle driving profiles can be incorporated into a simulated scene with a given road geometry, stationary objects, and moving vehicles. The occurrence of warning alerts and potential crash situations for this driving scene can then be analyzed with various radar sensor parameters and threat assessment algorithms. The warning system can be evaluated by varying the sensor and threat assessment parameters and the driving profile. This capability expands the ability to analyze crash scenarios beyond the real world test capabilities. It also allows a theoretical estimate of the impact of potential sensor redesign concepts on the collision warning algorithms.

The threat assessment algorithms are developed in C programming language independent of the simulation effort. A set of four algorithms has been integrated. The user display module is designed to allow easy integration with any number of algorithms as they are developed. To integrate those four algorithms into the display module, they are first translated into four functions, named alg1, alg2, alg3, and alg4. When the algorithms are supplied with proper inputs, they will return flags indicating the warning status. The user display module displays the top view of the driving scene and the warning status for each of the selected threat assessment algorithms. Users can choose from the menu which algorithms, in any, to use in the simulation and they can observe the warning status as the scene progresses with time.

The simulation tool is currently being used to evaluate threat assessment algorithms. First order effects can be observed and analyzed with the current simulation and model. This is a valuable analysis tool because crash situations are infrequent in the real world environment and are difficult to obtain meaningful data. However, further enhancements for the radar sensor model are required in order to perform a parameter variation analysis on the radar sensor parameters and the threat assessment algorithm parameters.

3.2.4 Threat Assessment Analysis

The purpose of threat assessment algorithms is to warn the driver with respect to potential crashes. The collision warning sensor tracks and identifies potential targets and passes this information to the threat assessment sub-system. The appropriate parameters for threat assessment can be range, range rate, following and target vehicle velocity, longitudinal and lateral accelerations, horizontal and vertical radius of curvature of road, target signature and quality, vehicle dynamics (such as yaw rate, heading angle, lateral position, steering angle, roll and pitch rates) and driver-vehicle interface parameters (windshield wipers, weather, road surface, brake status and powertrain status). On curved roads, multiple targets may generate false warning due to incorrect path determination, i.e., out-of-lane targets may generate false warning and in-lane-targets may be missed leading to a missed alert. The vehicle dynamics knowledge in combination with some sort of road geometry obtained by using vision-based lane sensing and GPS/map database is essential for determining the projected path of the vehicle. Ideally, very high angular resolution is required for multiple target tracking. Current systems with medium angular resolution may not be able to differentiate between 2 targets on a curved road, which could generate incorrect warning to the driver. Moreover, target's longitudinal and lateral accelerations that would predict trajectory are not available and can only be estimated roughly.

Since target acceleration is not available, current algorithms are based on fixed assumptions about target and following vehicle's decelerations for the rear-end collision scenarios. The two dominant scenarios are the inattentive driver and stopped object in the path. The collision warning algorithms are based on either computing the warning distance or computing the time to collision parameter or a combination of both approaches. Given the lead vehicle's velocity v_l, deceleration a_l, following vehicle's velocity v_f, deceleration a_f, range R (distance to lead vehicle) and the system delay time D (which includes the driver response time, brake response time and target acquisition time), the distance_based algorithms are given by:

Standard Driver Alert Equation

distsda = v_f² / 2a_f + D v_f - v_l² / 2a_l (3.1)

Closing Rate Equation

distcra = (v_f - v_l)² / 2a_f + D v_f (3.2)

If the computed values for distances (distsda or distcra) are greater than the range R, then a warning is issued to the driver.

The time-based algorithms are given by:

Time To Collision = R / (v_f - v_l) (3.3)

Time Headway = R / v_f (3.4)

A warning is given to the driver when the time to collision or time headway values exceeds some pre-specified threshold. The time-based warning algorithms are easy to evaluate but are incorrect for general collision warning scenarios, since no deceleration rate and system delay parameter is used. The distance-based algorithms also suffer from excessive false alarm alerts, as correct acceleration values are not being used.

The collision warning algorithms should warn the driver with minimum number of nuisance alerts and in sufficient time such that the driver can either avoid the crash or mitigate the crash. If the warning is either given too early or too often, then the driver will take that warning to be a false alarm or nuisance alert. A missed alert is defined as a warning which is either not given or is given too late for the driver to respond in proper time. In order to perform the sensitivity analysis, various crash and non-crash scenarios for lead and following vehicles with varying dynamics are set up for using simulated data. The scenario dynamics consist of a lead vehicle decelerating at constant rate and the following vehicle decelerating after a specified delay time. For a fixed set of parameters, analytical equations predict whether the type of scenario will be a crash or non-crash. The time to crash (for a crash scenario) and other vehicle dynamics parameters are also derived. The dynamics of both vehicles such as velocities, positions, decelerations and the collision warning algorithm's response are also computed.

Results of applying several algorithms (Equations 3.1- 3.4 with different parameters) on simulated data are given in Tables 3-3 through 3.5. The simulated data scenario was generated for lead and following vehicle velocities varying from 0 - 39 meters/sec and initial range varying from 5 - 150 meters. The number of scenarios for the simulated data set was 29,200. Using Monte Carlo simulation, the parameters for the lead deceleration, driver response time to lead vehicle braking and driver response time to alert warning were varied (sampled from distributions).

These simulations demonstrate that for an inattentive driver, the false alarm rate is very high due to the fixed set of acceleration/deceleration parameters being used. Moreover, false alarm rate can be minimized at the cost of hit rate. Ideally, the hit rate should be 100 % with the lowest possible false alarm rate. Table 3.6 shows Algorithm 1's performance on the data sets when true deceleration parameters are used. The corresponding results from Tables 3.3 - 3.5 are also shown as the first entry in each cell.

Preliminary analysis has shown that the false alarm rate for the inattentive driver can be reduced to less than 8 % when true deceleration parameters are used. The recommendation for future development is to evaluate algorithm performance and sensitivity when estimated acceleration/deceleration parameters are used.

3.2.5 Future Directions

The prototype sensor specified to the vendor met most of the requirements, however some critical parameters fell short of the specifications. The azimuth resolution is the most critical one of these parameters. Poor performance of the prototype radar with respect to this parameter results in clustering of two objects into a single object under certain conditions. For example, two vehicles in adjacent lanes are reported as a single object depending on other parameters, which was demonstrated by other tests. This may compromise the performance of forward collision warning system under certain conditions. The range resolution was another parameter that did not meet the specifications. However, this may not be as critical as the azimuth resolution since there may not be a need to discriminate objects as finely as stated in the forward collision warning application. The latency measured in the system can be minimized when the application is better integrated into the sensor, at this time it is not possible to quantify the components of this delay.

The most important issue is to be able to obtain unfiltered raw information from the sensor. Vendors do have embedded preprocessing optimized for a specific application. This of course optimizes the performance of the radar for a given application, however masks some important information that might be useful in evaluation of the sensor. The issue is not technical, most vendors consider the basic data as proprietary information. Another issue that was not investigated was the elevation field of view parameter. It was observed that the sensor detected many overpasses and overhanging road signs. The solution is to investigate both hardware and software techniques to come up with a cost-effective solution. For example, introducing a low cost but crude elevation angle resolution combined with signal processing techniques may improve the performance significantly. The vendors should place additional resources on improving their sensors working jointly with their customers.