Traffic Police Gesture Recognition For Automated Driving System

In recent years,benefits from the advancement of hardware technology and artificial intelligence technology,the research on automated driving system(ADS)is developing rapidly.It is one of the necessary conditions for the application of autonomous driving technology to recognize the traffic police gestures quickly and accurately and make correct response under complicated traffic conditions.Accordingly,the main theme of the thesis is the recognition of the gestures of traffic police for autonomous driving purpose.At present,the methods of obtaining traffic police gestures mainly include sensor-based recognition and vision-based recognition.This paper focuses on how to break through some key technical difficulties faced by traffic police gestures recognition based on traffic police video images captured by monocular camera and combined with effective deep learning algorithms.For example,factors such as complex situation at traffic intersections,poor lighting,gesture occlusion and too small traffic police characters in the field of vision(‘small people’)result in low gesture recognition accuracy,as well as complex models result in poor real-time performance.In this paper,using the cutting-edge methods in the field of computer vision for reference,skeleton sequence diagram and feature heat map are respectively used as the basis for gesture recognition,and the strategies for improving the recognition accuracy and efficiency of traffic police gestures recognition are explored.Specifically,this paper mainly carries out the following three research works:(1)Research on traffic police gestures recognition based on coordinate attention.For the low accuracy of traffic police gestures recognition caused by complex actual situation at traffic intersections and poor lighting,this study introduced Open Pose model to reduce the interference of lighting and other external factors to extract skeleton features,and introduced coordinate attention mechanism to further accelerate the recognition speed and enhance the generation accuracy of key points.Experimental results show that the proposed method have better recognition accuracy than similar algorithms,and has obvious advantages in the traffic police gestures data set containing weak light conditions.(2)Research on traffic police gestures recognition based on TransPose.For the possible partial occlusion of traffic police gestures in practical application,this study introduced the TransPose model that can predict the occluded key points.This model utilized the advantages of HRNet to maintain rich high-resolution representation and combined with the self-attention mechanism of Tansformer to explore the dependence between the key points of the traffic police’s body.Finally generate skeleton diagram to promote traffic police gestures recognition.Experimental results show that the proposed method has better recognition accuracy than similar algorithms,and has obvious advantages in the traffic police gestures data set containing occlusion.(3)Research on traffic police gestures recognition based on CenterGroup.In order to comprehensively deal with the low accuracy of traffic police gestures recognition caused by weak light,key points occlusion,‘small people’ and other problems,and to meet the requirements of running frame rate in practical applications,this study introduced a feature heat map based on CenterGroup.In this model,the high resolution network of Higher HRNet scale perception is firstly used to generate character center,and then the idea of Tranformer is introduced to solve the problem of key points occlusion.Experimental results show that the proposed method has better recognition accuracy than similar algorithms,and has obvious advantages in the traffic police gestures data set including weak light,key points occlusion and ‘small people’.The innovation of this study lies in :(1)Lightweight feature extraction backbone network based on coordinate attention mechanism was introduced to enhances key point features and effectively improves the traffic police gestures recognition accuracy under weak light conditions;(2)The TransPose model was introduced to predict the occluded key points through the dependence relationship between the key points,which effectively solved the problem of the occluded key points in part of the gesture;(3)Through the combination of center point positioning and self-attention mechanism to generate feature heat map,problems such as weak light,key points occlusion,‘small people’ and low frame rate in traffic police gestures recognition are comprehensively solved.