Pedestrian Detection And Tracking Based On Convolution Neural Network For Autonomous Driving

Autonomous vehicles are a hot topic in current social discussion and also the development trend of traffic in the future.Pedestrian detection and tracking is a very important part of driverless visual perception,which has a wide range of research significance and application value.Although many domestic and foreign universities,scientific research institutions and companies have made some research results in this area,but due to the complexity of unmanned driving scenes and the inherent characteristics of pedestrian characteristics,there are still many problems in pedestrian detection and tracking that need to be further studied.This thesis mainly studies the following three aspects:First,a real-time pedestrian detection algorithm based on improved SSD is proposed for the application of unmanned vehicles for pedestrian detection in real time.The traditional pedestrian detection method has low robustness and slow detection rate.Therefore,this algorithm adopts the one-stage target detection method SSD model based on convolution neural network with a higher detection speed.In order to improve the ability of the algorithm to extract target features under the real-time condition,the deep residual network with easy convergence and low model complexity was proposed as the basic backbone network of SSD model.Experimental results in the Caltech pedestrian dataset show that the algorithm has better detection performance and meets the basic requirements of pedestrian detection in driverless scenes.Second,aiming at the complex environmental background of driverless,as well as the large scale change and appearance difference of different pedestrians,the semantic segmentation features are gathered into the multi-scale target detection method based on one stage,so as to enhance the extraction ability of pedestrian features.At the same time,a network model of multi-feature joint learning is proposed without adding additional input.The experimental results on multiple pedestrian datasets show that the fusion algorithm proposed in this paper can significantly improve the accuracy of pedestrian detection and has a high detection rate.Finally,a Deep SORT pedestrian tracking algorithm based on fusion optical flow is proposed to solve the problem that the pedestrian detection method can not provide enough dynamic information of pedestrians,and the traditional online tracking method does not make full use of the characteristic information of pedestrians.In order to make full use of the characteristic information of pedestrians,the target motion information based on optical flow is integrated in the linear prediction part of Kalman filter.At the same time,in order to make the algorithm meet the real-time requirements,the improved SSD real-time pedestrian detection algorithm is taken as the target detection part of the Deep SORT model.The effectiveness of the method is shown on the dataset collected on the real campus.Finally,the results of pedestrian detection and tracking are mapped to the car body coordinate system through inverse perspective,which provides the directly available data for the decision and planning layer.