Traffic Police Gesture Recognition For Autonomous Driving

With the advent of the era of artificial intelligence,Driverless cars have become the mainstream trend in the development of intelligent transportation today.In order to be able to drive safely on the road,driverless cars need to make accurate judgments on complex road environments.When a driverless car encounters traffic congestion,signal failure,etc.,it is necessary to command the traffic through the traffic police.At this time,it is particularly important that the driverless vehicle can quickly and accurately recognize the traffic police gesture.Based on this,this paper studies the problems existing in traffic police gesture recognition.It mainly solves the problem that the traffic police gesture detection is difficult in the complex dynamic background,and the redundant motions in the traffic police gesture class and the small difference between the class motions interfere with the traffic police gesture recognition.The research content of this paper is mainly composed of the traffic police upper limb motion area detection algorithm,the traffic police key gesture extraction algorithm and the traffic police key gesture classification algorithm.The details are as follows:Firstly,aiming at the difficulty of traffic police gesture detection under complex dynamic background,an optimization algorithm based on LK optical flow method is proposed for the detection of upper limbs motion area of traffic police.The algorithm first introduces an image pyramid algorithm to compensate for the defect that the LK optical flow method is only suitable for small moving target detection.Then,based on the LK optical flow method,the concept of optical flow length and optical flow angle is proposed.Combined with the motion characteristics of the traffic police arm,the optical flow length and optical flow angle are used to select the feature points corresponding to the traffic police gesture,so as to extract the upper limb movement area of the traffic police.Finally,the 20 traffic police video samples collected were tested,and the detection results did not show any error.It provided an experimental basis for the extraction and identification of subsequent key gestures.Secondly,aiming at the problem of traffic police gestures,the key gesture extraction algorithm based on motion block is proposed.The algorithm first uses the image pyramid LK optical flow method to track the traffic police gesture and obtain the motion point.Then,using the techniques of graph clustering and entropy,the motion points are processed to obtain the motion blocks corresponding to the traffic police arm.Finally,combined with the periodic characteristics of the traffic police gesture,the motion trajectory of the motion block is analyzed to obtain key gestures.By comparing the key gestures obtained by the experiment with the standard key gestures,the key gestures extracted by the algorithm are basically consistent with the standard key gestures,and the effectiveness of the algorithm is verified.Finally,the transfer learning method is used to classify the traffic police key gesture pictures.For the problem of re-training convolutional neural networks on small data sets is likely to over-fitting or losing the slow convergence of the function.This paper firstly carries out data enhancement processing on the collected 2000 traffic police key gesture images.Using the VGG16 model,which performs well on the ImageNet dataset,performs transfer learning training on the enhanced data set,and the three transfer learning methods of the fine-tune method,only the top-level classifier and the re-training are experimentally compared,and the fine-tune transfer learning is obtained.The method is superior to the other two methods in recognition rate and performance,and the recognition rate reaches 97.73%.Finally,the data set before data enhancement and data enhancement is trained on the fine-tune transfer learning network,and the recognition rate after data enhancement is higher than that before data enhancement,thus verifying the effectiveness of data enhancement.