Recognizing Traffic Police Commanding Gestures Based On Computer Vision

In recent years,many cities have claimed that autonomous vehicles must obey the commanding gestures used by the traffic polices.Therefore,the needs for methods to recognize the traffic police gestures are urgent.Traditional methods for gesture recognitions rely on deep sensors or wearable sensors,which restrict their practical uses on road.Methods based on the vision sensors have less requirements on distance and lights,but due to the complexity of the vision information,analyzing and constructing the gesture models become difficult,it is hard for the hand-crafted feature extractors to recognize the images as accurate as human.There are two main challenges regarding the visual recognition of traffic gestures.The first one is that it's hard to extract skeletal data of traffic police from the image,due to the inconsistent practical situations and unpredictable factors such as occlusions,dressing changes.The second one it how to reduce or eliminate the interferences,such as human absolute positions and camera distances during the classification process.Leveraging the breakthrouth of deep learning in recent years,the pose estimation methods based on the visual information have developed a lot.This paper proposes a novel method which recognizes eight kinds of Chinese traffic police gestures based on visual information,and works at real-time.This method integrates the network of convolutional pose machine and hand-crafted features: bone length,angles with gravity as the spatial features,applies Long Short Term Memory to extract temporal features.The design of this method is balanced between accuracy and speed,provides the ability of recognizing traffic police gestures online.Meanwhile,this paper publishes a traffic police gesture dataset containing 2 hours of gesture videos and annotations,used for training and validation of our method.The dataset contains recorded videos with different environments and police dresses,the light conditions,distance between camera and human and complexity of background images also varied,which makes the experiment results reliable.In summary,this paper provides a vision-based method to recognize eight kinds of Chinese traffic police gestures and a gesture dataset with two hours of videos.It provides a solution for the autonomous vehicles to obey the traffic polices commands.