Research On Pedestrian Detection Technique In Advanced Driver Assistance System Based On Machine Vision

In recent years,the automobile industry has been rapid developed,car intelligentization level has gradually increased,and people have put forward higher and higher requirements for the active safety of automobiles.Advanced Driver Assistance Systems(ADAS)is the key to realize the active safety of automobiles,pedestrian detection is not only the technical difficulty of ADAS,but also the core technology of Pedestrian Collision Warning System,which is directly related to pedestrian safety.Due to the complexity of road scenes and pedestrians,there is no pedestrian detector with high detection accuracy,good real-time performance and strong robustness,at present,most pedestrian detection methods in automotive ADAS are based on Histogram of Gradient and Support Vector Machines,which requires manual design of pedestrian features,the steps are complicated and time consuming,and the deep learning visual algorithm provides a good solution,the feature extraction and feature recognition are unified into the same network model,and the pedestrian characteristics are automatically learned,which simplifies the detection steps and reduces the calculation.This thesis takes the automotive ADAS pedestrian detection problem as the research object based on the literature of related fields at home and abroad(Chapter 1),the following studies have been carried out:Firstly,a new hybrid dataset INRIA-NEW was established,and an end-to-end pedestrian detector was constructed by using deep learning visual algorithm Faster R-CNN or YOLO.The two key issues of pedestrian feature extraction and pedestrian feature recognition in pedestrian detection were studied in depth,and the extraction of pedestrian features and recognition of pedestrian features were unified into the same network model,and end-to-end pedestrian detection was completed(Chapter 2,Chapter 3).Secondly,a YOLO-Person pedestrian detector with high accuracy,good real-time performance and strong robustness was proposed.On the basis of YOLO,combined with batch normalization,residuals,feature pyramid network and pedestrian aspect ratio characteristics,the optimization scheme was established,and the migration learning method was used to train.And then,the YOLO-Person pedestrian detector was constructed,which achieved the balance of detection accuracy and real-time,and efficiently completed pedestrian detection(Chapter 4).Thirdly,the pedestrian collision warning criterion and corresponding workflow were proposed,a pedestrian collision warning system was built based on the optimization model YOLO-Person,and pedestrian detection in vehicle video was completed by YOLO-Person.The pedestrian collision warning system was modularized,and the corresponding solutions were given for pedestrian detection and pedestrian ranging.At the same time,the pedestrian detection test of the car video was carried out based on the optimization model YOLO-Person,the validity and superiority of the YOLO-Person method were verified(Chapter 5).Pedestrian detection experiments on INRIA-NEW hybrid dataset and vehicle video were carried out respectively in this thesis.The results show that the performance of end-to-end pedestrian detection model based on deep learning was significantly improved compared with traditional methods,and the YOLO-Person pedestrian detector constructed in this thesis has good performance in detection accuracy,real-time and robustness.