Research On Gesture Recognition Of Traffic Police Based On Deep Learning

With today’s major car companies and research institutions on the depth of learning technology research,auto-assisted driving technology is becoming more and more mature,many new vehicles are equipped with automatic driver-assisted driving function,the technology can help drivers to identify front and rear road conditions,intersection traffic lights,roadside traffic signs and other information and put forward corresponding driving recommendations.In the face of urban congestion or sudden situations on the road,there will usually be traffic police to the scene for traffic evacuation,and at this time,the autoassisted driving system needs to recognize the traffic police gesture signal,because this function directly affects driving safety,so the efficiency and accuracy of traffic police gesture recognition algorithm is particularly critical.Most of the current research on traffic police gesture signal recognition uses depth camera or multi-angle position to obtain traffic police skeleton model,while depth camera in the outdoor recognition effect is poor,detection distance is close and expensive,multi-angle position can not be mounted on the vehicle,so this paper proposes a single-eye two-dimensional color camera to obtain video sequence,and combined with deep learning algorithm to identify traffic police gesture signal.The specific contents of this paper are as follows:(1)Pre-processing the video sequence.In the actual traffic scene,due to air quality,natural light source,artificial light source,camera jitter and other factors interference,in the final acquisition of the film,more or less there will be some noise,in order to make the subsequent processing more smooth,better recognition,the use of digital image processing means will be obtained into the film for further processing,so that traffic police become clear,easy to identify.(2)The pre-processed video sequence is extracted from the traffic police bone joint point,and the bone joint point frame sequence is obtained.Aiming at the problem of slow and low accuracy of the traditional bone joint point feature extraction algorithm and the deep learning network framework such as VGG,ResNet and MobileNet,this paper puts forward the bone joint point extraction network composed of Hyperpose attitude recognition engine and the Tyny-VGG network structure based on VGG network,which greatly improves the extraction performance of bone joint points.The recognition network meets the requirement of extracting bone joint points in real time.(3)Time feature extraction and classification identification of the sequence of point frames of traffic police bone joints.In this paper,the coordinate sequence of traffic police bone joint points extracted by Hyperpose attitude recognition engine plus Tiny-VGG network is used as input to extract network input from traffic police gesture signal time characteristics.Time characteristic extraction network needs to use circulatory neural network,in view of the long-term dependence characteristics of traditional RNN network and the problem of large number of parameters and slow training speed of LSTM network,this paper puts forward the use of GRU gated loop unit as a time characteristic extraction network,effectively reduces the number of training parameters,improves the training speed and hardly affects the accuracy.The high-dimensional features extracted by the GRU network are then fed into the full link layer,and the probability of the gesture category is normalized through the Softmax function,wherein the output value is the largest as the final determination category.All the experiments in this paper are realized by Pytorch Deep Learning Framework,Hyperpose Attitude Recognition Engine,Pycharm Professional 2020.1.2 Integrated Development Environment,OpenCV Computer Vision Library combined with Python 3,C and programming languages,and finally the algorithm proposed in this paper is tested by using the open Chinese Traffic Police Command Gesture Data Set.