Pedestrian Detection Research Based On Deep Learning

Pedestrian detection refers to the detection of pedestrian categories in pictures or videos and the location of the target through computer vision technology.Thanks to the development of deep convolution neural network and target detection model,pedestrian detection technology is also more and more used in life,such as intelligent driving,security and robot.However,due to the diversity of pedestrian's posture,the complexity of the background,and the influence of light and angle,the task is very challenging.In this paper,the current mainstream target detection model is optimized based on the characteristics of pedestrians.Experiments show that the performance of the improved pedestrian detection and pedestrian instance segmentation model has been improved to some extent.The main research contents and innovations are as follows:(1)Build your own dataset.Although there are many existing public data sets,there are few for pedestrian instance segmentation.Therefore,1000 pictures were taken from life,including shopping malls,roads,parks,beaches and other places,covering different light in the morning,noon,night,and other living scenes such as cycling,children frolicking,three or five groups,followed by Pascal VOC and coco data sets are calibrated to get two self built pedestrian data sets,which are used to improve the data enhancement,training and testing of pedestrian detection and pedestrian instance segmentation tasks.(2)Improved Yolo V3 pedestrian detection model.Based on the model of Yolo V3,this paper proposes an improvement which is more suitable for pedestrian detection.First of all,according to the characteristics of self built data sets and pedestrians,K-means clustering is re used to generate a new anchor size;secondly,in order to better play the network performance,the residual connection block in the backbone network Darknet is re designed by using the idea of group convolution,which improves the accuracy of the model without adding network parameters and complexity.The experimental results show that the improved network performance has been improved to a certain extent.(3)Improved block RCNN pedestrian segmentation model.Based on the mask RCNN model,an improved pedestrian segmentation algorithm is proposed.Aiming at the deficiency of FPN module in mask RCNN in using semantic information of feature channel,cfpn module is proposed and applied in mask branch.In this module,high-level features are expanded by means of up sampling and fused with low-level features.New feature groups are specially used for subsequent instance segmentation tasks.Finally,in the comparative experiments of various backbone networks,it shows that the improved pedestrian instance segmentation model has made great improvement in ap50 and map.