Intensive Pedestrian Detection Algorithm Optimized With Exclusion Loss

In pedestrian detection,the detection accuracy which is not high enough and the difficulty of NMS threshold setting are necessary to solve.For the problem of insufficient detection accuracy,there are many mainstream object detection networks trying to improve it,such as RCNN series,SSD series,YOLO series,etc.Although the detection accuracy of the object is continuously improved with the improvement of the network structure,the final detection accuracy still has a lot of room for improvement.In another way of thinking,post-processing can also improve the accuracy of target detection,such as NMS algorithm,Soft-NMS algorithm,etc.However,for NMS algorithm,it is easy to have difficulty setting threshold,if two objects are close,if the threshold is set too high,it is possible to retain a false positive that should not occur.If the threshold is set too low,it is possible to delete the nearby detection boxes,so that the detection network does not have an ideal detection effect on such a scene.In order to solve this problem,Soft-NMS has been proposed.This method mitigates the problem by weighting the score of the final detection boxes,but a threshold still needs to preset in the process of deleting the redundant detection boxes according to the score.However,how to set this threshold is still uncertain,so Soft-NMS still has room for improvement.In order to improve the above two problems,this paper takes the high-speed rail monitoring pedestrian object detection as the background,and starts from loss function of the detection network and the detection network structure,and proposes a method combining the exclusion loss and the pedestrian recognition model to solve them together.The main innovations of this paper are as follows:(1)Introduce the exclusion loss in the detection model for training.This method can make the detection boxes obtained by the detection network more concentrated in the ground-truth corresponding to itself,and away from the other ground-truths and the corresponding prediction boxes,so that When the NMS threshold can be set in a loose space,it can effectively reduce a large number of false positives and improve the difficulty of setting the NMS threshold.(2)Design a network with strong classification ability to further process the pedestriandetection boxes obtained before to determine whether it is really a pedestrian.The prediction boxes of the pedestrian is retained,and the prediction boxes of the background is suppressed,thereby further eliminate the false positives and improve the final detection accuracy.(3)Combining the above two methods,we get our final pedestrian detection model,which has both advantages above.Although it will increase the overall running time by a small amount compared with the original detection model,it will increase the detection time by a small amount.It is acceptable to improve the detection AP of the entire detection model,which is meaningful for practical applications.