| In recent years,driving assistance system and unmanned driving technology have gradually emerged and become research hotspots.Among them,traffic police command gesture recognition as a key technology,has been widely concerned by researchers.In particular,the research of real-time gesture recognition algorithm for traffic police command is of great value.The video data not only carries the spatial dimension information of the static image,but also the characteristic information of the time dimension.Therefore,the spatial and temporal information in the video data can be extracted by the fusion of interframe motion features and single frame static features,which can be used for traffic police command gesture recognition.This paper designs a real-time recognition algorithm of traffic police command gestures based on the fusion of interframe motion features and multiple features.The main research contents are as follows:(1)To solve the problem that there is no open source database in the field of gesture recognition of traffic police command,combining the characteristics of gesture signal of traffic police command and the task requirements of gesture recognition of traffic police command,a set of Gesture video data of traffic police command,including edited video and unedited video,is established to support the subsequent research on gesture recognition algorithms of traffic police.(2)Aiming at the problems of high similarity and difficulty in distinguishing traffic police command gesture signals from different perspectives and non-standard traffic police command gesture signals in practical application,this paper selects the Optical Flow guided Feature(OFF)to describe the inter-frame motion and describe the inter-frame motion information from the time dimension.Static features such as color and texture are extracted from a single frame of static image,and information such as characters and scenes are described from the spatial dimension to improve the ability of the model to distinguish different command gesture signals.In addition,in order to improve the modeling ability of the model for long-term motion information,this paper designed the OFF-TSN network model,and then used the features of the whole video clip to recognize the command gesture of the traffic police.The experimental results show that the OFF-TSN network model can effectively extract spatio-temporal features for the traffic police command gesture recognition task at the video clip level and can be used for recognition.The model achieves a recognition rate of 93.4% and a recognition speed of 206 FPS in the test on the self-built traffic police command gesture video clip data set,which has good performance.(3)Based on the OFF-TSN network model,this paper designs the real-time recognition algorithm of traffic police command gesture.In order to solve the problem of delay and missed detection caused by large amount of data brought by multi-scale sliding window sampling,this paper adopts target detection algorithm based on SSD to detect the presence of traffic police in real-time video and control the start and end time of sampling.A filtering network based on the Off-TSN network model is designed to filter the sample fragments lacking gesture information.In order to solve the problem of location and recognition of command gesture in real-time input video,a location recognition network based on the OFF-TSN network model was designed.By improving the loss function of the OFF-TSN network model,the network could judge the integrity of command gesture signal.Finally,the location recognition network is combined with the detection sampling algorithm to realize the location and recognition of command gestures.In the test on the self-built unedited video data set of command gestures,the normalized Levenshtein distance between the recognition result and the real label is 0.872,and the recognition speed is 26.8 FPS,thus realizing the real-time recognition of command gestures of traffic police. |
PDF Full Text Request