Research On Pedestrian Detection In Urban Traffic Scene Based On Key Point Network

As an important direction of target detection,pedestrian detection technology is widely used in intelligent monitoring(IVS),humancomputer interaction(HCI),auxiliary driving(ADAS)and so on.However,due to the interference factors such as multi-pose,crowded occlusion and complex background,pedestrian detection algorithm still has low recognition rate and high missed detection rate,and can not meet the real-time problem in some specific scenarios.Therefore,how to improve the accuracy of pedestrian detection model while achieving realtime detection effect is a very important issue.Based on the pedestrian detection network of the central key points,this paper studies the pedestrian detection in urban traffic scenes from the perspective of improving the accuracy and real-time performance of the algorithm.The main research work is as follows:(1)Aiming at the problem of serious occlusion of pedestrians in urban traffic scenes,this paper proposes CSANet based on dual attention mechanism based on CSP,which can well realize the integration of pedestrian channel features and spatial features,and adds deep empty volumes to the feature extraction network based on ResNet-50 to improve the extraction ability of the model for pedestrian high semantic features.The DIoU-NMS is introduced to suppress the pedestrian prediction box under occlusion more reasonably,which further improves the detection ability of the model for occluded pedestrians.The MR-2 of CSANet in Caltech is reduced to 4.69%,which is about 17.9%higher than that of CSP.(2)Due to the urban traffic background,the multi-scale problem of pedestrians is more obvious.Therefore,this paper proposes a bidirectional feature pyramid detection algorithm,which adds bidirectional FPN on the basis of CSP to fuse the pedestrian semantic information of the deep feature map and the pedestrian position information of the shallow feature map,and optimizes it through network pruning to reduce the calculation amount.Finally DFNet achieved MR-2 of 5.06%on Catlech.(3)In this paper,MR-2 is used as the evaluation index of the detection system,and CSANet is cascaded into a complete detection system CSDFNet.In addition,in order to improve the speed of the model,lightweight network is introduced for acceleration.Experimental evaluation was carried out on Caltech and CityPersons of the dataset.Finally,the overall model’s pedestrian detection performance for severe occlusion is 19.1%and 22.5%higher than that of ALFNet and O-RCNN,respectively.The performance of multi-scale pedestrian detection is 24%higher than CSP,and the average detection time per image is 19 ms.According to the experimental results of CSDFNet on Caltech and Cityperson datasets,this detection algorithm improves the accuracy of pedestrian detection and meets the requirements of real-time.