Research On Pedestrian Detection Algorithm Based On On-board Video

With the development of artificial intelligence,people put forward higher requirements for the intelligence of cars,especially for active safety.Pedestrian detection,as one of the key technologies of active safety,directly affects the lives of pedestrians and is an important technology to reduce pedestrian casualties.At present,pedestrian detection based on traditional machine learning have complex steps and poor robustness.The detection effect is poor in complex road environments.Although the pedestrian detection algorithm based on deep learning can learn pedestrian characteristics autonomously and has high detection accuracy,most of the algorithm models require a high amount of computation,which is difficult to meet the real-time demand in the process of high-speed driving.Therefore,this paper takes the pedestrian detection of vehicular video as the research object,expecting to find a pedestrian detection plan that can meet the real-time demand,and has good detection accuracy and robustness.On the basis of reviewing relevant literatures at home and abroad,the following researches are mainly carried out.(1)Firstly,based on the analysis of the shortcomings of traditional image sensors,a dynamic vision sensor with low redundancy,little influence from light changes and high acquisition frequency is proposed for image acquisition.It can reduce the redundancy of pedestrian picture and improve the real-time of detection system.Then,the algorithm principle of YOLO series is described and compared with traditional pedestrian detection algorithm,RCNN series algorithm and SSD series algorithm.Through the performance analysis of the algorithm,YOLOv3 with high frame rate and good detection accuracy is adopted as the algorithm framework of pedestrian detection.(2)Secondly,a camera based on dynamic vision sensor is used for image acquisition to build a new pedestrian data set.A wavelet threshold denoise method based on improved genetic algorithm and generalized cross validation is proposed to solve the problem of large number of noise points in dynamic vision sensor imaging.Through the improved genetic algorithm,a new fitness function was established to obtain a better threshold of denoise and improve the desiccating effect of the wavelet threshold algorithm.Through experiments,it is found that the YOLOv3 algorithm has higher detection accuracy on the picture after drying.(3)Next,an improved YOLOv3 algorithm is proposed for the striking feature and less feature in the image of dynamic vision sensor.K-means ++ algorithm is adopted to cluster the new pedestrian data set,and six priori box sizes more in line with pedestrian attitude are obtained to improve the detection accuracy.Then,the ordinary convolutional layer in YOLOv3 network is changed to the depthwise separable convolutions,so as to achieve the purpose of compression and acceleration.Remove the Darknet53 structure used by YOLOv3,design a more streamlined feature extraction layer,reduce the final detection scale,reduce the computation required by the algorithm model,and improve the real-time performance of the algorithm.(4)Finally,an experimental platform was built to train the network based on the mainstream deep learning framework Caffe,and training parameters were adjusted to obtain a better IM-YOLOv3 algorithm model.The performance evaluation of the algorithm model is carried out on the collected data set,which proves that the algorithm has good real-time performance and improves the detection effect to some extent.Finally,experiments are carried out on several kinds of scenes that are difficult to detect,and the algorithm has a good detection effect in these cases.It provides a new scheme for the pedestrian detection of on-board video.