Research On Traffic Police Gesture Recognition Method Based On Deep Learning

For intelligent driving vehicles to achieve real automatic driving,accurate and fast recognition of the traffic police gesture is essential,so the realization of the recognition of the traffic police gesture is of great significance to promote the automatic driving technology.For the recognition of traffic police gestures,the traditional methods mostly adopt the method of wearing sensors or depth sensors to collect the motion characteristics of traffic police gestures,and then classify them according to the different motion characteristics of each gesture.Although these two methods realize the recognition of traffic police gestures to a certain extent,they cannot be popularized in a large scale.With the popularity of deep learning,this technology has been applied in more and more fields.Therefore,this article adopts deep learning method to solve the problems that pedestrians and traffic police are difficult to distinguish and the gesture recognition of traffic police is low.The main tasks are as follows:1.Realize the detection of traffic police.The first step in realizing the recognition of traffic police gesture is to distinguish traffic police from pedestrians.The biggest visual difference between traffic police and pedestrians is that they wear reflective clothing.Therefore,from the perspective of reflective clothing of traffic police,this article adopts the technology of target detection in deep learning to distinguish traffic police from pedestrians through yolov5 network model.The experimental results show that this method can realize the traffic police detection and ensure the real-time performance of the algorithm.2.Static traffic police gesture detection.For the traffic police gesture,each gesture has a key gesture frame to represent this gesture.Therefore,combined with the practical application,this article improved the method of replacing the convolution layer,adding the attention module and replacing the loss function of yolov5 network model,and applied the improved algorithm to the recognition of traffic police gestures.Experimental results show that the accuracy of the improved algorithm is improved by 3.39% compared with that of the unimproved algorithm,which has a good application value.3.Recognition of dynamic traffic police gestures.In order to enable the algorithm to realize the recognition of dynamic traffic police gestures,this article adopts the method of combining spatial temporal graph convolutional network based on multi-attention and spatial temporal long and short-term memory network to realize the recognition of traffic police gestures.The structure attention mechanism and time attention mechanism were added to the spatial temporal graph convolution network model to improve the algorithm’s extraction of the topological structure of the skeleton points and the motion features of independent nodes.It is combined with the spatial temporal long and short-term memory network model to improve the ability to extract the periodic motion features of traffic police gestures.Experimental results show that the accuracy of the combined algorithm model is improved by 2.7% on average compared with the original algorithm model,which has good research value.There are 42 figures,3 tables and 77 references.