Predicting Driver Intent and Maneuvers for Road Safety

本研究计划的重点在于识别影响交通安全的驾驶认知因素，定义自动车辆安全技术设计的合理原则，以及开发智能高级驾驶辅助系统（i-ADAS），以驾驶员行为预测和纠正为安全改进的核心原则。更准确地说，本研究的目的是发展对驾驶动作的操作理解，这可能用于未来i-ADAS实现中预测引擎的概念。为此，我们设计了一辆实验车辆，该车辆能够在感知到的车辆3D正面环境中以绝对坐标记录其即时正面环境、车辆里程计、驾驶员操作和车辆功能操作、驾驶员头眼行为（头部姿势和3D凝视）以及3D凝视点（PoG）。我们从16名司机那里获得了超过3TB的数据，这些司机沿着安大略省伦敦市的预定路径行驶。设计了自动注释该数据集的算法。例如，在标记过程中采用了多车道检测、车辆检测、地平面检测和gps校正技术。我们现在正在设计一种技术，使用与时空相干性相关的约束在语义上分割3D立体数据流，以获得物体描述符及其3D边界体。反过来，驾驶员的3D凝视可能与这些边界体相交，从而识别驾驶员正在看什么，以及在哪里。我们感兴趣的是在0.25到2秒的时间窗口内预测最可能的驾驶员机动：当预测的机动与车辆和驾驶员所处的当前交通状况不一致时，这个时间足以让i-ADAS执行缓解措施，并避免更大时间窗口带来的预测可靠性问题。最近用我们的数据和使用逻辑回归的3层神经网络进行的实验表明，对于代表大约一小时驾驶的数据序列，在接下来的0.5到1秒内，标准机动加速和减速可以以99.6%的精度预测。其他典型的演习也曾以这种精确度被预测过。我们目前正在开发和测试并行方法来解决驾驶员机动预测问题。我们非常有兴趣确定在我们的驾驶序列中哪些数据流对机动预测影响最大（有证据表明，在这方面头眼行为很重要）。

The focus of this research proposal rests on the identification of cognitive factors involved in driving that impact traffic safety, the definition of sound principles for the design of automated vehicular safety technologies, and the development of intelligent, Advanced Driving Assistance Systems (i-ADAS), with driver behaviour prediction and correction as the central tenet of safety improvement. More precisely, the objective of this research is to develop an operational understanding of driving maneuvers that may be used in the conception of prediction engines incorporated into future implementations of i-ADAS. Toward this end, we instrumented an experimental vehicle capable of recording its immediate frontal environment in 3D, the vehicle odometry, driver maneuvers and operations of vehicular functions, driver cephalo-ocular behaviour (head pose and 3D gaze), and the 3D Point of Gaze (PoG) in absolute coordinates within the perceived 3D frontal environment of the vehicle. We proceeded to obtain more than 3TB of data from 16 drivers, on a predefined path around the city of London, Ontario. Algorithms to automatically annotate this data set were devised. For instance, novel techniques for multi-lane detection, vehicle detection, ground-plane detection, and GPS-correcting techniques were employed in the labeling process. We are now devising techniques to semantically segment the 3D stereo data stream using constraints related to spatiotemporal coherence, in order to obtain object descriptors along with their 3D bounding volumes. In turn, the 3D gaze of drivers may be intersected with these bounding volumes, allowing for the identification of what the driver is looking at, in addition to where. Our interest is in predicting the most probable driver maneuver in a time window from 0.25 to about 2 seconds: this amount of time is sufficient for an i-ADAS to perform mitigating actions in cases when the predicted maneuver is inconsistent with the current traffic situation the vehicle and driver find themselves in, and avoids the prediction reliability problem posed by larger time windows. Experiments recently conducted with our data and a 3-layer neural network using logistic regression show that for a data sequence representing approximately one hour of driving the canonical maneuvers accelerate and decelerate can be predicted with an accuracy of 99.6 percent for the next 0.5 to 1 second. Other canonical maneuvers have been predicted with this level of accuracy. We are currently developing and testing concurrent approaches to the problem of driver maneuver prediction. We are very interested in determining what data streams in our driving sequences have the most impact on maneuver prediction (there is evidence that cephalo-ocular behaviour is important in this regard).