基于深度学习的交警动态手势检测与识别方法研究

Research on Traffic Police Dynamic Gesture Detection and Recognition Method Based on Deep Learning

文中基于深度学习方法设计了交警动态手势检测与识别算法,以Top-down的姿态估计方法获取交警人物关键点建立人体骨架图,采用时空图卷积的方式进行动作识别.设计交警目标检测算法、交警姿态估计算法和交警动态手势识别算法,对输入模型的骨架图设计了尺寸归一化,令算法对不同尺寸的骨架图具有相同的识别性能,提高了算法的鲁棒性.所设计的方法在限制容许错误率在10%、15%、25%以及50%的条件下能够达到最高96.32%的识别率.

Based on the deep learning method,a dynamic gesture detection and recognition algorithm for traffic police was designed.The key points of traffic police characters were obtained by the Top-down attitude estimation method to establish the human skeleton diagram,and then the action recognition was carried out by the convolution method of space-time diagram.By designing traffic police target detection algorithm,traffic police attitude estimation algorithm and traffic police dynamic gesture recognition algorithm,the skeleton diagram of the input model was designed with size normalization.The results show that the designed algorithm has the same recognition performance for skeleton images of different sizes,which improves the robustness of the algorithm.The designed method can achieve the highest recognition rate of 96.32%under the condition of limiting the allowable error rates to 10%,15%,25%and 50%.