Road Vehicles And Pedestrian Detection Based On On-board Video

With the rapid development of world economy and significantly enhance of the quality of life,it is widely used motor vehicle to attend social activities every day,which leads to the social harmful phenomenon occurring frequently,such as environmental pollution,road congestion and traffic accidents.In order to ease and solve the relevant problems,some scholars from all over the world are committed to researching the intelligent transportation system.Meanwhile,with the enormous improvement of computer hardware?s computing power,it is no longer a research obstacle that makes the real-time computation of some complex algorithms.As an important part of unmanned auxiliary system,the development of vehicle and pedestrian detection technology has a great practical significance.In the technology of vehicle and pedestrian,it is difficult to use one single method of feature extraction and classification to two kinds of target detection,because both of them have many differences in their target characteristics.But,target detection scheme based on deep learning through calculate of complex neural network for multiple targets can effectively get the characteristics collection of the detection and classification,so as to get the favor of international scholars.In this article,in order to be satisfied with the real-time detection requirement of the road vehicle and pedestrian based on on-board video,we put forward using deep learning regression network YOLO for vehicle and pedestrian detection,and designed the corresponding data set for training and testing the model.Then we join with Deep-SORT algorithm as the core method of multi-object real-time tracking to track the detected targets for a period of time,which can overcome the weakness of YOLO that ignored the connection information between up and down in the video when making target detection.The introduction of the tracking algorithm,which greatly relieves the phenomenon of target “lose” when using YOLO to have target detection based on on-boaed video,at the same time,it is also has inhibition of target occlusion.Due to the use of the target detection and tracking in the process of target detection based on on-board video,two methods maybe mark a same target,which leads to target duplicate detection,and makes the lower performance of detection.To curb the above situation,the Hungarian algorithm is used in the Deep-SORT to make data correlation between detection box and tracking box,and remove the repeating box areas.And for design the elements of data correlation matrix,this article makes a weighted fusion scheme between the Markov distance for movement matching and the minimum cosine distance for apparent matching based on feature vector,then get the corresponding numerical information to matching of Hungary.In this article,we are detecting the lane line in front of the train vehicle based on related technology of image processing.Algorithm is mainly composed of image pre-processing,Gaussian filtering to de-noising,canny edge detection,hough transform to detect line and lane line detection based on k-means clustering algorithm method.At the same time,the result of the lane line cooperated with the output of vehicle and pedestrian detection as the final test result.With the algorithm is adopted in this paper,the experimental results show that the algorithm is feasible and effective.Compared with the target detection algorithm based on YOLO,the algorithm of our article improves the target detection rate of vehicle and pedestrian based on keeping a real-time detection standard.