基于Kinect的地铁乘客不安全行为识别方法与实验

Kinect-based recognition method and experiments on unsafe behavior of subway passengers

为解决地铁视频监控技术对乘客不安全行为只记录不识别且较少考虑识别精确度的问题,提出1种基于Kinect传感器的高效识别方法。以Pelvis为向量起点和动作活动高频关节为终点构建识别特征向量;运用余弦定理获得标准动作与测试动作关节的最大角度差序列;以最大角度差为动作特征量建立相似度计算模型,运用动态时间规整算法(DTW)将初始结果转换为动作相似度。以相连关节法为对照组开展对比实验,结果表明:前者在抽烟、挥拳、挥手呼救等行为识别的准确度分别为91.7%,86.9%,89.2%,平均比对照组高4%以上,显著提高了地铁乘客不安全行为的识别率,可为地铁智能安全管控提供理论与技术依据。

In order to solve the problem that the subway video surveillance technology only records but does not recognize the unsafe behavior of passengers and considers the recognition accuracy less,an efficient recognition method based on Kinect sensor was proposed.The characteristic vectors of recognition were constructed by taking Pelvis as the starting point of the vector and the high-frequency joints of the motive action as the end point,and the maximum angle difference sequence between the standard motion and the test motion joint was obtained through the law of cosines.A similarity calculation model with the maximum angle difference as the motion characteristic quantity was established,then the dynamic time warping algorithm(DTW)was used to convert the initial results into the motion similarity,and the comparative experiments were carried out with the connected joint method as the control group.The results showed that the accuracies of the former on behavior recognition such as smoking,punching,and calling for help were 91.7%,86.9%,and 89.2%,respectively,which were on average 4%higher than the control group.It significantly improves the recognition rate of the unsafe behaviors of subway passengers,and can provide theoretical and technical basis for the intelligent safety management and control of subway.