夜晚车辆异常事件分析

夜晚车道模型是车辆跟踪和车辆行为分析的基础，但是当高速公路或者城市道路光线较暗时，很难通过车道检测的方法来建立车道模型，夜晚车辆快速行驶或相邻帧车辆之间重叠度较低时无法实现准确跟踪。针对此类问题提出了一种基于学习的车道模型建立方法和基于多帧的最佳匹配跟踪方法。首先利用自动多阈值分割方法提取场景中光亮的目标；其次，利用车灯的相关特征移除非车灯光亮区域；接着，利用空间信息把车灯聚类成一个车辆目标，利用多帧的最佳匹配跟踪方法进行跟踪；最后利用车辆跟踪参数与车道模型的融合对夜晚车辆异常事件进行分析。实验结果表明，该算法能够准确地检测出夜晚车辆换道、逆向行驶、交通拥挤、停车等异常事件，并且有很强的鲁棒性。

Vehicle abnormal events analysis at night

The vehicle lane model is the base of vehicle tracking and vehicle behavior analysis. However, it is difficult to establish vehicle lane model through lane detection algorithm because it is dark on the highway or urban road, and it is difficult to track vehicle exactly when the video frame rate is slow or the vehicle’s speed is too fast. Therefore, the learning based vehicle lane model establishment method and the multi frames based best matching tracking method are proposed. Firstly, a fast bright object segmentation process based on automatic multilevel histogram threshold is applied to extract bright objects effectively. Secondly, some no vehicle bright regions are removed by some features of vehicle's lamps. What’s more, the lamps are clustered into a car object by using spatial information and tracked by the multi frame based best matching tracking method. Finally, tracking information and vehicle lane mode are used to analyze abnormal events. Experimental results show that the algorithm can exactly detect abnormal events such as changing lane, reverse driving, heavy traffic, stopping car etc at night and it has strong robustness.