RoIs Extraction Method Research And System Implementation For Vehicle-mounted Thermal Imaging Pedestrian Detection

With the increase of automobiles in China,traffic congestion and traffic accidents are becoming more and more serious.The advanced driver assistance system based on the vehicle-mounted thermal imaging pedestrian detection can provide driver with danger warning,which is an important way to prevent traffic accidents.The infrared camera is suitable for night scenes,however,as the thermal imaging image lacks texture,it is more difficult to identify the target in the image.The vehicle-mounted thermal imaging pedestrian detection system running on embedded system is limited in computing resources.However,the system is strict in the accuracy of the detection algorithm and the instantaneity of the system.RoIs extraction is an important part of the pedestrian detection system,which can reduce the object search range in the images effectively.Thereby it can improve the calculation efficiency and improve the detection accuracy of the system.The adaptive dual-threshold segmentation method is a RoIs generation method with high precision of RoIs but very low computational complexity.It is very suitable for the vehicle-mounted thermal imaging pedestrian detection system.Based on the benchmark pedestrian detection framework of “RoIs extraction-SVM classification”,this paper systematically studies the related technologies of RoIs extraction.The main research contents include image enhancement scheme of thermal image,optimization of threshold segmentation algorithm for thermal image,and RoIs fine filtering method based on human component characteristics.The main contributions of the paper are as follows: 1)Design an expanded maximum filter to enhance the image.By the filter the brightness of the pedestrian target can be increased,and pixels proportion in image can be expanded.2)Study the adaptive double-threshold segmentation algorithm,and and based on the analyses a double-threshold segmentation algorithm enhanced by Haar-like operator is proposed.Besise,a fast implementation scheme of the algorithm is given.The new algorithm can not only improve the quality of RoIs,but also improve the calculation efficiency.3)The cascaded RoIs filters are proposed.They ares used to filter RoIs to reduce interference to classifier.We intergrate three components proposed in the paper into the pedestrian detectionsystem and test the system on our public thermal imaging pedestrian detection data set.According to the testing result,the pedestrian detection accuracy is improved and the system computational cost is reduced on the premise of taking the recall rate into account.