Recognition Of Traffic Command Gestures Based On Deep Learning

In recent years,with the development of software and hardware technology,autonomous driving technology has become a popular research topic.As an integral component of autonomous driving technology,the recognition of traffic command gestures is increasingly crucial.According to the regulations recently released by the police department in Beijing,autonomous vehicles must be capable of recognizing traffic command gestures.Therefore,the recognition of traffic command gestures has become an important direction in the research of autonomous driving technology.Traditionally,recognition methods of traffic command gestures rely on devices such as wearable sensors and depth sensors.However,due to the inherent limitations in these methods,they cannot meet the requirements of autonomous driving technology.With the significant breakthroughs of deep learning technology in the field of computer vision,there have been tremendous advancements in the recognition of traffic command gestures based on vision sensors,making it one of the most promising solutions currently.The recognition of eight traffic command gestures is presented in this thesis from two aspects,namely static images and dynamic videos based on human body keypoints.In order to improve the accuracy of the algorithm,two algorithms is optimized in this thesis.The optimized algorithm can meet the accuracy of the intelligent assisted driving system.The main contributions of this thesis are summarized as follows:(1)Traffic command gesture images is captured in this thesis and is expanded to 16000 images using nine data augmentation methods.The 16000 traffic command gesture images are labeled using the label Img software.In this thesis,an annotation interface is utilized to annotate60 traffic command gesture videos frame by frame for the video datasets.(2)A traffic command gesture recognition method is proposed in this thesis based on static images.Firstly,the experimental environment required by the YOLOv5 algorithm is built using Anaconda.Then,the YOLOv5 algorithm is used to train the traffic command gesture image dataset,achieving an average accuracy of 83.3%.The YOLOv5 algorithm is modified by incorporating an attention mechanism module into it,due to factors such as the green background,in this thesis.The average accuracy of recognizing traffic command gestures is improved by 2.5% compared to the YOLOv5 algorithm without the attention mechanism module.Finally,the ORBBEC camera on the ROS operating system is used in this thesis and the YOLOv5＿ROS algorithm with the added attention mechanism module is called to recognize traffic command gestures.(3)Traffic command gestures from dynamic videos are aimed to be recognized in this thesis,based on human body key points.Firstly,a convolutional neural network is utilized to extract 14 key points of the human body.Then,Traffic command gesture features,including normalized skeletal length and the angle between the skeletal gravity direction,are extracted.The traffic command gesture recognition based on dynamic video is realized through the traffic command gesture recognition memory network(Gate Recurrent Unit).Finally,a comparison is made with existing methods,and the accuracy is improved by 1.1%.At present,the traffic command gesture recognition technology based on vision sensors has become an essential feature in intelligent assisted driving systems.This technology not only alleviates urban road traffic pressure but also improves the driving experience and enhances driving safety.Reliable application value is provided for the intelligent assisted driving system to recognize traffic command gestures by the two methods proposed in this thesis.