Research On The Pedestrian Intrusion Identification In Autonomous Driving Scene Through Machine Learning

Pedestrian-vehicle accidents highly likely result in incapacitating injuries and fatalities.Pedestrians’ behavior is extremely difficult to predict,they can change their moving direction in a very short period of time,and are prone to traffic safety accidents.Because of the lack of necessary protection measures,they are often fatally injured in traffic accidents,so the death rate of pedestrians has bee increasing sharply in the latest years.To avoid such accident,computer vision based efforts have been made in driving scene understanding and pedestrian protections.Accurate estimation of pedestrian’s motion are very helpful to make reasonable decisions and reduce pedestrian mortality.However,at present,most research uses single algorithm to characterize pedestrians’ behavior in a complex driving scene,it is easy unable to get data because of occlusion.For this problem,in this work,we proposed a pedestrian intrusion identification method based on pedestrian detection and pedestrian tracking methods.The details are as follows:(1)Aiming at the accuracy and real-time requirements of pedestrian detection in the automatic driving scene,this paper employed the algorithm of pedestrian detection based on yolov3,and based on the Keras deep learning framework.The video in the real traffic scene is taken by the car camera as the experimental data.The pedestrian detection accuracy of yolov3 on the self built dataset is～96%,and the average frame rate(FPS)is about 28 frames per second.The results show that the accuracy and speed of pedestrian detection meet the actual requirements.(2)To solve the problem that it can’t get the object relying on a single feature to track pedestrians when encountering occlusion,this paper proposes a tracking method combining Kalman filter and optical flow method.Yolov3 can accurately and real-time detect the pedestrian and output the bounding box.Kalman filter tracks the center of the bounding box to obtain the pedestrian position information in real time.At the same time,it uses a kind of neural network(TVNet)which can understand the characteristics of optical flow from the data to obtain the pedestrian speed information.The results show that the combination of yolov3,Kalman filter and optical flow method is more accurate than the tracking method relying on a single feature,which greatly enhances the accuracy and robustness of pedestrian tracking.(3)This paper mainly studies the single pedestrian when the vehicle is still or moving slowly.Firstly,yolov3 and Kalman filterare integrated into a framework to extract the characteristics of pedestrian trajectory and displacement,and then the optical flow method is used to extract the characteristics of pedestrian speed.On the basis of extracting multi-dimensional feature vectors,we proposed a pedestrian intrusion identification method based on support vector machine.The experimental results show that the accuracy of this method is 95.45% while the recall is equally 93.33%,which proves that the proposed method is highly effective in pedestrian intrusion identification,and can provide decision-making basis for the vehicle in the automatic driving scene.