Research Of Traffic Police Gesture Recognition Method Based On Pose Estimation

In the rapid development of the present,people's living standard is gradually improved,the number of vehicles is also increasing,the number of drivers through the driving test is also increasing.Since it is rare for traffic police to direct the traffic,most drivers cannot fully recognize all the gestures of traffic police as time goes by,which will cause inconvenience to traffic police in directing the traffic.In the auxiliary driving system,the traffic police gesture recognition function will provide great convenience to the driver.At present,the rapid development of communication technology,based on 5G or even 6G communication Internet of vehicles technology will be widely used,vehicles can directly through the network to obtain traffic signals,access more convenient.However,in special cases,there may be poor network signal or no network service.Based on the minimum security consideration,the vehicle itself also needs to have the ability of information collection and judgment.In the near future,driverless cars will be ready for commercial use,and driverless technology will become more and more mature.The accurate perception of the surrounding environment is a necessary condition for the unmanned driving technology.Even in the era of unmanned driving,there may be emergencies.In the case of network and machine failure,the traffic police need to direct the traffic at the scene,so the recognition of the gestures of the traffic police is a necessary ability of the unmanned driving system.The current gesture recognition technology of traffic police adopts the method of computer vision,which is affected by illumination,background and other factors,and the recognition accuracy is low,so there is still considerable room for development.In this paper,the gesture recognition of traffic police is carried out based on the possible problems in practical application and attitude estimation algorithm.The specific research content of this paper is as follows:(1)A traffic police gesture recognition method based on ROI region detection,human skeleton key point extraction and variable length sequence recognition is proposed.In the past gesture recognition of traffic police,ROI region detection of traffic police is usually not included.In this way,the network not only needs the ability of gesture recognition,but also needs the ability of character recognition of traffic police.Thus,the difficulty of training is increased imperceptibly,resulting in the decline of recognition accuracy.The scheme proposed in this paper recognizes the traffic police in the first step,and then only focuses on the gesture recognition itself.(2)The gesture estimation algorithm was used to replace the Kinect sensor used in previous studies to extract key features of human bones,which was more simple and convenient in hardware.Based on CPM(Convolutional Pose those)to improve the network structure,make it more suitable for single bone extract key points.(3)The variable length sequence gesture classifier is proposed,and the fixed length sequence recognition is changed to the variable length sequence recognition,which is more consistent with the characteristics of different gestures in the traffic police gestures with different time lengths and more suitable for the real scene.In addition,the method of feature template is adopted to judge whether the traffic police gesture to be recognized matches the template gesture according to the similarity network.Different from the previous gesture recognition method,the traffic police gesture is directly classified,which increases the extensibility of the overall solution.