One of the challenges of lane departure warning (LDW) systems is to differentiate between normal lane keeping behavior and lane change events in which drivers simply do not use the lane change indicator. Lane keeping behavior differs between drivers and often between driving scenarios, therefore a static threshold of predicting steering maneuver is not an ideal solution. The objective of the current study is to develop an adaptive method of predicting driver lane change maneuver using vehicle kinematic data.
The paper presents an adaptive steering maneuver detection algorithm, which can detect the earliest indication of driver’s intent to change lanes. The overall approach was to observe the driver’s “normal” lane keeping behavior for a period of time, and seek driver lane keeping behavior which falls outside of what is “normal” for each specific event. We modeled normal driving behavior in this study using a bivariate normal distribution to continuously monitor the vehicle distance to lane boundary (DTLB) and lateral velocity measured in most production LDW systems.
The results of our algorithm were validated against visual inspections of 949 randomly selected lane change events from the 100-Car Naturalistic Driving Study (NDS), in which we compared the time of driver steering initiation estimated by the algorithm against visual inspection. The comparison between algorithm results and visual inspection shows that all steering initiation in lane change events in the sample occurred within 5 seconds of lane crossing. In addition, a sensitivity analysis on the bivariate normal distribution boundary shows that the contour line representing 95% probability produced the lowest average percentage error (2%) with an average delay of 0.7 seconds between the algorithm predicted driver steering initiation time and video inspection. The resultant algorithm was deployed in a large subset of 100-Car and was able to identify the steering initiation time in a total of 53,615 lane change events. The resultant algorithm shows utility in assisting future active safety system in monitoring driver lane keeping behavior, as well as providing active safety system designers further understanding of driver action in lane change maneuvers to improve designs of LDW systems.