Research On Video Human Detection Method

Video pedestrian detection is the basis of video pedestrian behaviors, poses processing, which is responsible for determine the existence and getting the precise position of human target in the video. This paper mainly do two aspects of the work: The training framework and the features of human detection algorithm based on statistical learning and cascade classifier detection framework gets the optimization and improvement. At the same time, this paper adjusts the Fast RCNN network structure of the detection model and proposes a multi scale Fast RCNN network structure that is more useful to video human detection according to the characteristics of video human detection.Human detection based on the statistic learning and cascade classifier detection framework mainly finished three works:Firstly, sampling the large scale data sets through positive bootstrap and training the classifier with small data sets instead of large data sets, which make the original cascade classifier deal with the large scale data sets effectively and even salve the problem that the statistical learning algorithm depend on samples, but can’t deal them effectively. At the same time, Reducing the negative sample bootstrap range, compressing negative sample bootstrap time and improving training efficiency through negative sample inheritance. Secondly, using the Uniform LBP feature to train a pre filter classifier and filter the candidate target. By the end, the speed of detection get improve through reducing the candidate target range and eliminating the bad influence of high complexity of COHOG features on detection speed. Lastly, improving the detection accuracy by combining the detection and tracking that could reduce the miss of human detectionHuman detection based on multi scale Fast RCNN network structure mainly deal with the problem that the Fast RCNN network structure is difficult to balance the human target detection at different scales because of video human large scale change. This paper adjusts the framework and trains a network with branches, the different branch of which deals with different scale human. In the finally, fusing the human body target detection to get the last detection result.This paper uses the standard datasets test the performance of them. Experimental results show that the two algorithms have achieved nice results.