Research On Guiding Action Recognition Method For Unmanned Vehicle

With the development of urbanization in China,unmanned vehicles are gradually becoming popular.In the face of the coming era of unmanned vehicles,various countries have long established standards for unmanned vehicles on the road,which stipulate that unmanned vehicles must be able to recognize traffic signals and traffic gestures,and must give priority to obeying traffic police command.However,complex backgrounds,different lighting and shuttling vehicles and pedestrians have a great impact on the interference of gesture recognition in the actual traffic gesture recognition process.To address the problem of low accuracy of dynamic gesture recognition in complex background situations in guiding action recognition for unmanned vehicle,the research focuses on the recognition of the guide,real-time guidance action recognition and other techniques,and the main research contents are as follows.(1)To address the current problem of low accuracy of single recurrent neural network model in action recognition,this thesis proposes a combined LSTM-GRU network for action recognition.First,the network principles and structures of LSTM and GRU are analyzed,and the respective advantages and disadvantages of LSTM and GRU are discussed.The best way to assign weights is found for the output of LSTM and GRU modules during the training of the combined network,and then the combined training is performed.The experimental results show that the average accuracy of the proposed combined LSTM-GRU network on the traffic police gesture dataset for eight different traffic police gestures is improved compared with the accuracy of a single LSTM or GRU.(2)To address the problem of low accuracy of guided action recognition under complex road conditions,this thesis proposes an optimization method based on Blaze Pose guided action feature extraction.First,the accuracy and real-time performance of different human pose recognition networks are experimentally compared,and finally Blaze Pose is adopted as the pose node recognition network.Secondly,the characteristics of the guide’s movements are analyzed,the whole-body features are improved,and a spatial feature extraction method focusing on the upper body limb movements is proposed.Finally,the classification and recognition of the movements are performed by the combined LSTM-GRU network.The experimental results show that the proposed guided action extraction method in this paper has improved the accuracy and real-time performance of guided action recognition compared with previous methods.(3)To address the problem that there will be pedestrian interference during the guiding action recognition of the guide,this thesis designs and implements an unmanned vehicle guiding action recognition system.The face recognition module is added to the guidance action recognition algorithm,so that only the designated guide will do effective guidance action,and others will be meaningless action.The system mainly contains three modules,image acquisition module,guide identification module and guiding action recognition module.And the system optimization scheme is proposed for the system speed degradation problem.The method and system designed in this thesis satisfies the task of recognizing and classifying the real-time guidance actions of designated guides in complex background situations,while ensuring real-time performance.