Research On Multi View Traffic Police Gesture Recognition Based On Composite Model

The progress of science and technology has accelerated the development of driverless technology.Nowadays,many types of driverless vehicles have been running on real roads.Although driverless vehicles bring great convenience to people when they travel,the problems that follow cannot be ignored.For example,when there is an emergency on the road,the traffic lights and traffic police at the intersection may command the traffic at the same time.In this case,the priority of traffic police command is higher than the traffic lights.If the driverless vehicle only responds to the signal of the traffic light and ignores the command of the traffic police,it may cause unnecessary traffic accidents.To solve this problem,this dissertation improves the target detection algorithm and time series model,and designs a parallel operation architecture to achieve real-time and dynamic recognition of traffic police gestures on different directions of lanes.The main research contents of this dissertation are as follows:(1)The problem of traffic police area extraction in the screen.In the pictures collected on the real road,in addition to the traffic police,there may also be other characters.In order to eliminate interference,this dissertation compiles a web crawler to crawl a certain number of images containing traffic policemen and pedestrians from the image website to create a data set to train the target detection model algorithm,screen the characters in the screen,and select the screen containing only traffic police,so as to facilitate subsequent operations.(2)Dynamic gesture recognition.To solve this problem,this dissertation uses Google’s open source framework Mediapipe to extract key human points,and encapsulates them into groups according to a certain number.After that,the time series model is used to train the encapsulated gesture feature data and make a template,which can adapt to the data based on time series to realize the final dynamic gesture recognition.(3)Hand gesture recognition of traffic policemen from different perspectives.In reality,for the same traffic police gesture,different interpretations can be obtained by observing it from the front and the side.To solve this problem,this dissertation determines the recognition orientation of the traffic police gesture by detecting the face orientation of the traffic police,and trains different models of the gestures corresponding to different orientations,so as to achieve the effect of recognizing the traffic police gesture from multiple perspectives.(4)Optimization of composite model performance.This dissertation designs a parallel architecture,distributes target detection,improved LSTM and Fast DTW to different processes,and uses shared variables between processes to store the necessary data for inter process communication,so as to improve the execution efficiency.