Design And Implementation Of Vehicle Infrared Pedestrian Detection System

When drivers are driving at night,due to the exhausted physical conditions,complex light source,unclear sight and low pedestrian brightness,the pedestrians in front of vehicle are not easy to be noticed.Hence,traffic accidents at night are more prone to happen than it during the daytime.However,accurate and efficient vehicle infrared pedestrian detection system can instantaneously find pedestrians in front of the road and simultaneously remind drivers to avoid potential accidents,which has very important practical significance in improving traffic security.Therefore,the main task of this paper concentrates on the design and implementation of the vehicle infrared pedestrian detection system.This paper carries on the deep and comprehensive research in the following three aspects in order to achieve the accurate and efficient vehicle infrared pedestrian detection system.Firstly,the infrared pedestrian detection algorithm.Focusing on the Otsu's method(referring to maximum interclass variance threshold segmentation algorithm)and aiming at the problem of segmentation failure in the infrared scene with the small proportion of pedestrian targets,this paper presents an OTSU threshold segmentation algorithm under the hotspot detection,which is combining the characteristics of infrared scenes and the analysis of OTSU algorithm.This algorithm can effectively solve the threshold segmentation failure problem of the traditional OTSU and therefore,enhance the background segmentation effects.On the other hand,according to the miscellaneous and time-consuming shortcomings of tag positioning in the Region of Interests(ROIs)of the traditional infrared pedestrians,this paper proposes a ROIs fast tag positioning algorithm with the improved infrared pedestrian characteristics analysis.Through the design of ROIs marking rules,regional merging,noise filtering and other operations,this algorithm achieves ROIs rapid marking positioning of infrared pedestrian in the O(n)time complexity,and the time complexity is reduced to 1/23 of the traditional method's time complexity.Secondly,the realization of vehicle hardware processing system.According to the requirements of the vehicle infrared detection system,this paper designed the hardware processing system based on RK3288 + Linux-3.10,which solved the hardware design shortcomings of the current vehicle infrared pedestrian detection system.Meanwhile,under the analysis of pedestrian characteristics,this paper proposes an excellent pedestrian detection algorithm: Deformable Parts Model(DPM)and Latent Support Vector Machine(Latent SVM)based on directional gradient histogram feature.Furthermore,this paper extends this algorithm into the infrared scene and uses it as a pedestrian detection algorithm within the infrared pedestrian detection system.Thirdly,the optimization and acceleration of the infrared pedestrian detection algorithm.Under the designed hardware processing system and the improved infrared pedestrian ROIs extraction,combining the methods of loop expansion,function inline,compiling optimization,parallel computing etc.,this paper improves the performance of DPM and Latent SVM algorithm in terms of code,compilation,and program structure.Based on the constructed infrared pedestrian sample library,the infrared pedestrian classifier is trained to realize the infrared pedestrian detection.Experiments show that the optimized DPM and Latent SVM algorithm can reach 91.36% accuracy of the infrared pedestrian detection.And the hardware system can achieve the detection rate of 23 frames per second with1/17 of time-consuming before optimization,which basically satisfy the requirement of accurate and real-time vehicle infrared pedestrian detection.