基于3D骨架的交警指挥姿势动作识别仿真

Recognition Simulation on Traffic Police's Command Postures Based on 3D Framework

在智能交通指挥、交警教学培训等领域中,需要对交警指挥动作进行识别和准确度评估。针对现有系统对交警指挥动作描述不合理,对复杂场景识别率低的问题,提出了一种对交警指挥姿势动作的识别方法,并据此设计了一套虚拟交警指挥动作训练系统。首先利用人体深度图像构建人体三维骨架,采用人体骨架特定关节点的相对空间距离对单个姿势进行描述。然后对原始动作姿势序列进行关键姿势帧提取,并采用DTW算法对八种交警指挥动作的关键姿势序列进行训练和识别。最后将识别结果映射到某虚拟十字路口场景中进行交警指挥动作训练。实验结果表明,改进方法平均识别率为98.875%,具有较强的鲁棒性,对低光照和复杂场景适应性也比较良好。

In fields of Intelligent Transportation System and traffic police training courses, recognization and accu- racy rating evaluation of traffic police＇ s command postures are needed. Aiming at the irrational description of traffic police＇ s commanding postures in the current system and the low recognition rate of complex situations, a method for recognizing traffic police＇ s commanding postures was put forward in the paper, and a set of virtual training system was also designed based on the method. Firstly, we build a 3D body skeleton with human body range image, and de- scribed each posture by adopting the relative space distances between specific joints of human skeleton. And then, we conducted key postures frame extraction to the original behavior and postures sequence, and adopted the DTW method to train and recognize the key posture sequences of the eight kinds of traffic police＇ s commanding postures. Finally, we mapped the recognized results to virtual crossroads to train the traffic police on the commanding postures. Experiment results show that the average recognition rate of this method can reach up to 98. 975%. The method is of strong robustness and can adapt well to low illumination conditions and complex background conditions.