A Pedestrian Detection Method Based On Convolutional Neural Network

With the rapid development of social economy,people are strongly demanding intelligent life.As an important part,pedestrian detection affects people's life all the time.In recent years,deep learning has risen rapidly in the computer vision,which can be used in many fields such as recognition,detection,segmentation and so on.And it also provides a new approach to pedestrian detection task.Pedestrian detection is to lo-cate all pedestrians in images or videos using rectangular boxes with confidence scores.In practice,pedestrian detection suffers lots of problems such as poor image quality,different appearances and postures,occlusion,and various sizes of detection targets.Therefore,the deep learning method is applied to this study to detect the pedestrians.And the contribution of this dissertation includes as follows:1)We propose a method named DC-CSP to enhance pedestrian detection by com-bining pedestrian density and score refinement,which includes C-CSP,D-CSP and DC-CSP.Firstly,to increase the confidence scores of pedestrians and simultaneously de-crease that of background in the final decision,we design a classifier module on the basis of the CSP subnetwork to form C-CSP method.Besides,a stage score fusion(SSF)rule is also designed to update the detection scores by utilizing the complemen-tarity of the detection head and classifier modules.Secondly,for reducing the missing of occluded pedestrians and false detection of a single pedestrian,a density estima-tion module is added to CSP subnetwork to form D-CSP method.According to the estimated pedestrian density map,an improved adaptive NMS(IAN)post-processing method is also proposed to further improve the detection results.Specifically,a high IoU threshold would be used for mutually occluded pedestrians in order to reduce miss-ing detection,while a low IoU threshold would be used for a single pedestrian in order to reduce false detection.Thirdly,combine all of above modules and rules to form our DC-CSP method.For enhancing the feature extraction ability of the network,we use hourglass network as the backbone and conduct relevant experiments again.Compared with other existing methods,our DC-CSP method achieves the best performance on multiple benchmark datasets.2)For the sake of solving other problems in pedestrian detection,we also propose a method named AF-ALFNet by combining attention mechanism and feature fusion,which includes A-ALFNet,F-ALFNet and AF-ALFNet.Firstly,aiming at reducing the missing of pedestrians occluded by other objects,we propose a method named A-ALFNet which is combined with attention mechanism.Data enhancement can increase the number of occluded pedestrians,while attention module can focus on learning their features.Secondly,it is found that small pedestrians are often missing in detection,we apply feature fusion to ALFNet subnetwork to form F-ALFNet.The overall network can make use of context information to obtain more accurate feature maps,reducing the missing of smaller pedestrians.Thirdly,combine all of above modules to form our AF-ALFNet method.In order to achieve better detection results,we fuse the location and classification information of DC-CSP and AF-ALFNet.The experimental results demonstrate the effectiveness of our method,which can improve 2%to 3%on each subset of Citypersons dataset.In this dissertation,we propose two different pedestrian detection methods to al-leviate some issues of pedestrian detection,such as missing detection,false detection,and inaccurate confidence scores.The above two methods will be widely used in our future life.