Research Based On Deep Learning For Crowded Pedestrian Detection

With the continuous development of machine learning and computer science,using deep learning method to detect pedestrian has become the mainstream,and gradually began to be applied in safe cities,intelligent transportation and other fields.However,pedestrian objects in crowded scene have inner-class occlusion and large-scale changes,which seriously restrict the accuracy of classical deep learning detection methods such as SSD,Faster RCNN and so on.Therefore,we study the crowded pedestrian detection method based on deep learning,and improve the method from multi-scale feature fusion,multi-task detection and asymptotic location.It is of great significance to improve the accuracy and adaptability of pedestrian object detection under occlusion and scale change conditions.Our main work can be summarized as follows:Firstly,we propose a stability weighted multi-scale feature fusion method(SW-MFF).For the weighted sum of the five original feature maps with different resolutions extracted from the basic network and the five deep feature maps upsampled from the final feature maps,for the latter has more abstract and stable feature expression ability,we increase the weight of the latter in the fusion process.Then,we construct the multi-scale convolution feature,and detect the region of interest.Experiments on HUST self-made datasets and CUHK public datasets show that compared with Faster RCNN,the miss rates of our method on two datasets are reduced by 3.4% and 2.6% respectively(FPPI=0.1).Secondly,we propose an object classification,bounding-box regression,and drift depression-based multi task detection method(OBD-MTD).In addition to classification and regression tasks,aiming at the serious occlusion of crowded pedestrian objects in class,we utilize a loss function to model the overlap degree between the predict windows and the adjacent occlusion objects.The purpose of drift depression is achieved by minimizing the loss function,which is applied to both RPN and RCNN stages in detection network.Experiments using SW-MFF as benchmark method show that the miss rates on HUST and CUHK datasets decrease by 2.3% and 1.1% respectively after adding multi-task detection method(FPPI=0.1),which has better accuracy and robustness for occlusion change.Thirdly,we propose a cascaded RCNN-based asymptotic location model(CR-ALM).On the basis of two amendments to the predict windows in RPN and RCNN stages,we cascade the second-stage RCNN module to modify the predict windows several times,and gradually generate a more accurate set of predict windows.Based on this,an asymptotic locationing model is constructed.Finally,the model is integrated with the first two methods of this paper.Experiments using SW-MFF + OBD-MTD as benchmark method show that CR-ALM method can generate more accurate predict windows,effectively eliminate redundancy and false predict windows,and the miss rates on HUST and CUHK datasets are reduced by 2.2% and 1.2%(FPPI=0.1),respectively.