| With the development of politics and economy,countries,enterprises and individuals all over the world pay more attention to the security cause.The monitoring system has developed from the original analog video with manual monitoring to the middle stage of the semi digital storage and then to the all digital monitoring system of today.In the monitoring system,the intelligent front-end monitoring system is outstanding,the diversity of the needs of the algorithm and the rapid development of the front-end processor,making it widely used.Over the years,the researches of pedestrian detection has continued.Pedestrian detection algorithm plays an important role in the application of electronic bayonet,pedestrian avoidance system in unmanned vehicle,passenger flow detection and so on.Intelligent front-end with the function of pedestrian detection can detect targets and analyze the video stream,replacing manual monitoring and reducing the resources of transmission bandwidth and storage.In this paper,the function and performance requirements of the intelligent front-end system for pedestrian detection are designed according to the technical requirements of the real time intelligent analysis equipment for security surveillance video.Determine the network camera with DM8127 as the hardware platform of the system after analyzing the advantages of DM8127.Analyze the five modules of pedestrian detection intelligent front-end,and considering the software and hardware platform of the system,choose a method which combine the Support Vector Machine(SVM)with Histogram of Oriented Gradient(HOG)features based on the previous implementation of the pedestrian detection system.Simulate the pedestrian detection system on MATLAB,including the extraction of HOG feature module,the image pyramid detection module and the multi window fusion module.According to the real-time and accuracy requirements of pedestrian detection in intelligent front-end,The corresponding solutions are given for the three characteristics of HOG features:1)The scaling invariance of HOG features.Select the same size image containing different height of pedestrian as test data,design three layer pyramid based on former pyramid image detection principle.Conclude that the pedestrian height in the range of 88 pixels-128 pixels can be detected by window size of 64*128.So this paper gives a single image pyramid detection method based on this conclusion.2)The high dimension of HOG feature.The high dimension of feature dimension,which leads to the long time to extract features and detection.parameters can be adjusted properly on the premise of ensuring the accuracy.3)The HOG feature is less effective for occluded pedestrians.Pick pedestrians only with the upper body as part of the positive samples.According to the coordinates given by detection system display pedestrian box in the interface,and in the principle of multi-dimension window fusion technology,this paper gives a recursive fusion algorithm.After simulation,the LIBSVM is transplanted to DM8127 for pedestrian identification,and the algorithm of extractiong HOG features is transplanted to realize the whole intelligent monitoring system.And then test the whole system.The results showed that the addition of pedestrians contains upper body as positive sample can solve the problem of occlusion of pedestrian effectively.On the premise of accuracy,adjust parameter when extracting HOG feature can improve the speed of detection.The given recursive fusion algorithm can work effectively.And transplantation of HOG combined with SVM to DM8127 can detect more than 90%pedestrians. |
