Research On Vehicle Detection Based On Faster RCNN

Vehicle detect using vision is a vast important research subject of object detect based on vision,which should locate and classify the latent vehicle region of still images or video frame by some algorithms.Sadly,the system for detecting vehicle and pedestrian based on vision still can’t deal with the real transport scene,especially the ones that the background diverse dramatically,although decades of effort have been spent on research of it.Meanwhile,the development of deep learning has advanced at a tremendous pace,such as the deep convolution neural network apply on image classification,object detection,semantic segmentation and so on.Since it has surpassed traditional algorithms,pedestrian detection or vehicle detection based on it gradually became new fashion of automatic technology and intelligent traffic system,which has been pay high attention to by industry and academia.This paper has carried out the application of the system of pedestrian and vehicle detect based on deep convolution neural network with open source database as the research object.This paper conducted the following research contents,(1)This paper summarized the existing object detect algorithms,which was split to traditional ones and ones based on deep convolution neural network,especially the principle of Faster RCNN.(2)Designed a K-Means cluster algorithm like method used for generating parameters of anchor on top of on the statistical result of the height attribute of the vehicle tag.For each ground truth,our algorithm just takes the nearest 4 anchors as count for computing the area’s intersection-over-union(IOU)between ground truth and its assigned anchors.(3)Given the refine comparison of residual network and the inception structure of GoogLeNet,this paper made full use of residual learning structure and the vision attention mechanism for reconciling different scale feature maps.(4)Modified traditional algorithm of example mining by two terms,in which the first one is taking the top N classification loss of negative samples for back propagation,and the another one is transforming the predicting parameters of location of positive samples to image coordinate to compute the area’s IOU between each positive sample and its assigned ground truth for weighting the location loss of positive samples.(5)This paper tried the best to train the model with different hyper parameters,and chose a best one according the result in test database.This paper make full comparison between the original algorithm and different advanced module attached to the baseline.This article has the following three main results: Firstly,this paper thoroughly analyzed the principle of faster RCNN and its shortcoming when applied on pedestrian detect.Secondly this article designed some useful perfection based on the analysis above and completed the training work and comparison work,with the results witnessing the effectiveness of model.Finally,this article makes inclusion of the point for advancing fortunately.