Pedestrian Target Detection Method Based On Deep Learning

Pedestrian detection is an important part of intelligent monitoring.Detecting pedestrians in surveillance video automatically can drastically reduce the waste of human resources.Pedestrian detection is more difficult than general object detection because of pedestrian's non-rigid characteristic,occlusion,complicated background and other issues.In this paper,Faster R-CNN used for general object is applied to pedestrians in the surveillance scene,and the problems that appear in the algorithm are optimized.Detection model of general object to detect pedestrians in surveillance scenes will cause more false positives,because the training samples of the classifier aren't representative.The hard example mining to generate representative samples can enhance classifier and solve false positives.The existing pedestrian detection algorithm use the heuristic of hard example mining that only label the samples around GT.The location information can't guarantee representativeness of samples.So,we propose a novel method to mine representative negative samples,which comprehensively using the confidence and location information of candidate regions.This method doesn't need any extra hyper-parameters.Therefore,it doesn't increase the training complexity and computation of Faster R-CNN and is easy to implement.With the deepening of the model,the features are more advanced and can effectively describe the pedestrian.However,some pedestrians have less available information in the surveillance scene due to shooting angle,distance and occlusion.The more layers there are,the less fine-grained information the feature has.So,the detection is less effective.In this article,we use multi-layer feature fusion to increase the fine-grained information.In addition,we have adopted Faster R-CNN's shared-convolutional thinking for real-time detection by changing previous feature fusion.Therefore,we not only improve the detection results but also ensure the real-time performance.Experiments on standard databases(INRIA)and databases in the surveillance scene(PKU-SVD-B and Caltech)are performed in this paper.The performance indicators are better than the standard Faster R-CNN.This shows that the method proposed in this paper can effectively solve the above problems.