Research On Traffic Police Command Gesture Recognition With Artificial Features And Deep Learning

When traffic jams,accidents or signal lights are encountered in road traffic control,a traffic policeman is required to command traffic by means of gestures.Therefore,an unmanned motor vehicle needs to be able to quickly and accurately recognize traffic police command gestures.Deep learning can effectively extract common features from the input high-dimensional information through sample learning to achieve classification.Its application makes video-based traffic police command recognition performance greatly improved.However,deep learning requires large sample data and the training process is long.The recognition speed and accuracy of existing methods cannot meet the real-time and reliability requirements of unmanned driving.The obvious features extracted by prior knowledge are still valuable,such as the skeleton features closely related to human posture,and the characteristics of the video light stream reflecting the motion.This paper studies the combination of valuable artificial features and deep learning for automatic identification of traffic police command gestures,and improves in-depth network structure and algorithms,effectively improving training speed,recognition speed,and recognition accuracy.The specific research content of this article includes:(1)The objection of traffic police in complex environments and the judgment of manual command traffic conditions.In order to solve the problem that the traditional method requires manual extraction of features and experience,and the problem of poor recognition stability,this paper proposes a traffic police objection method based on SSD and a traffic command situation judgment method based on average background modeling and human aspect ratio.Some simulated samples are collected.Experiments were conducted on the actual field collection sample data set.The results of the recognition did not appear to be missed.The average detection time per frame was 0.0196 s.The traffic policeman could accurately locate the traffic police through the data enhancement technology.It is the premise of subsequent traffic police command gesture video segmentation and recognition.(2)Traffic police command gesture video segmentation.The traffic police command gesture video data often contains multiple consecutive command gesture actions and needs to be segmented into a single gesture action video in order to further identify what kind of gesture.This paper proposes a fast segmentation of continuous motion video based on human aspect ratio and movement rate of change.This method can quickly determine the starting frame and ending frame of traffic police command gesture.(3)Research on the classification and identification method of traffic police command video combined with artificial characteristics and deep learning,specifically combined with skeleton,optical flow information and video input directly from the video to the depth network for traffic police command gesture recognition.Because the shallow 3DCNN has not been able to well characterize traffic police command gestures,this paper uses deep learning algorithms based on C3 D and ConvLSTM to realize three-channel traffic police command gesture recognition combined with artificial features and deep learning.The advantage of this method lies in the use of Skeleton data can effectively avoid background interference and optical flow can capture the characteristics of the motion information,while the problem of the difference is solved through deep network structure and deep learning,and ultimately accurately identify the traffic police command gestures.Finally,simulation experiments show that this method can effectively improve the recognition accuracy of traffic police command gestures,and the average recognition rate on the collected eight traffic police command gesture data sets can reach 97.87%.Next,with the increase in the number of training samples,the system performance There will be further improvement.This method can be applied to applications with obvious artificial features and high input signal dimensions.