Research Of Pedestrian Detection Algorithm Based On Image

Pedestrian detection is one of the hottest topic in computer vision area.Detection algorithm got a rapid progress with the development of computer hardware among the past several decades.However,the effect degraded when it comes to real life occasion.Our paper dedicated to the research of pedestrian detection approach,improved the performance of traditional region proposal algorithm and the pedestrian detection convolution neural network.The main contribution of our article has two aspects.First,a coarse to fine region proposal framework has been put forward to refine the bounding boxes which generated by traditional region proposal method.All generated boxes would be rescored by a Bayes Prediction Model which trained with several geometrical features in the proposed framework.Experimental results demonstrate that the detection average precision increase at least 15% when applying to pedestrian.Second,our paper improved the effect of deep learning based detection network Faster R-CNN.During the region proposal stage,multi-layer region proposal approach was used to improve the scale problem in pedestrian detection;Before RoI Pooling Layer,more convolutional feature layers were merged to enhance the descriptive ability of final feature;At last detection stage,multi-region loss function was suggested to detect more occluded pedestrian.Experiments shows a 10% average detection precision increase when comparing with the original Faster R-CNN detection network.