Development of the methods for tracking and behavior understanding of vehicles and pedestrians in urban traffic scenes

上一财年，我们开发了一种既能跟踪图像中的刚体目标，又能跟踪图像中的行人等柔性目标的算法。本财年，我们对该跟踪算法进行了改进，使我们能够获得这些类型的运动目标的详细轨迹，并实现了一个实时系统的原型。另一方面，我们开发了一个基于实验的事件检测系统，该系统主要是基于算法的每个对象的时空轨迹，该系统利用基于规则的推理方法对交通事件进行分类。首先，坐标类算子从跟踪结果中提取每辆车的时空轨迹；第二，事件检测系统行为类算子从时空轨迹中提取每辆车的“熄火”、“停车”等行为；最后，事件类算子利用图像上车辆的行为信息来检测交通事故、拥堵等交通事件。如上所述，事件检测系统采用了由三种算子(坐标类、行为类和事件类)组成的语义层次，与人类的认知过程具有很高的亲和力，可以说，车辆和行人同时跟踪的算法的开发取得了一定的积极成功。在事件检测方面，重点研究了车辆行为识别的基本算法。我们认为该算法同样适用于车辆和行人等图像，并将继续对系统的可扩展性进行验证。

Last fiscal year, we developed an algorithm which is able to track both rigid objects such as vehicles and flexible objects such as pedestrians in images.This fiscal year, improving this tracking algorithm, we enabled to get detail trajectory of these various types of moving objects, and implemented a prototype of Real-time System.On the other hand, we developed a system which detects events based on experiments mainly conducted on Metropolitan Expressway.By applying the Spatio-Temporal trajectory of each object from the algorithm, this system classifies traffic events by using a rule-based reasoning method.First Coordinate-class operator extracts Spatio-Temporal trajectory of each vehicle from tracking results.Secondly, Behavior-class operator extracts behavior of each vehicle such as "Stall" and "Stop" from the Spatio-Temporal trajectories.Lastly, using behaviors information of vehicles on the image, Event-class operator detects traffic events such as accidents, congestion, etc.As above, our event detection, system utilizes a semantic hierarchy consisting of three kinds of operators (Coordinate-class, Behavior-class and Event-class) and has high affinity with the human cognitive process.So, it can be said that development of algorithm which tracks vehicles and pedestrians simultaneously achieved some positive success. When it came to the event detection, fundamental algorithm was developed to recognize behaviors of vehicles in particular. We consider that this algorithm can also be applied to images including vehicles and pedestrians, and will continue the verification of the system extensibility.