Research On Improved YOLO Pedestrian Detection Algorithm

In recent years,pedestrian detection technology has been widely used in the fields of intelligent monitoring,intelligent driving and intelligent robots.Pedestrian detection technology refers to the technology that locates the pedestrian target in the specified image through the algorithm and selects the specific location and range,and plays a pre-processing role for the subsequent work.However,the pedestrian targets in pictures and videos have different shooting angles,different pedestrian postures and certain occlusions,resulting in insufficient feature extraction ability of the YOLOv3 algorithm for such targets;at the same time,the YOLOv3 algorithm is insufficient for the feature information fusion of pedestrian targets.All these problems bring certain difficulties to pedestrian detection.In response to the above problems,an improved YOLOv3 algorithm is proposed to improve the accuracy of pedestrian detection.The main research work includes:（1）Aiming at the problem that the YOLOv3 feature extraction network has insufficient extraction ability,which makes some pedestrian targets undetectable,an improved ECA-Res2Net-YOLO model is proposed.First,the hierarchical residual structure Res2Net is used to replace the DBL structure of the original YOLOv3 feature extraction network.While reducing the model complexity,the multi-scale feature extraction ability is enhanced.Second,an efficient channel attention mechanism ECA is introduced to capture the dependencies of adjacent channel features and enhance the network’s attention to pedestrian targets.Finally,we emb efficient channel attention ECA after 1×1 convolution in the Res2Net module to re-associate the features after grouping and channel splicing.The ECA module re-weights the channel weights of the features extracted by Res2Net,in order to focus on useful pedestrian features and suppress useless features.At the same time,the feature extraction network of YOLOv3 is improved based on this module.The experimental results on the VOC pedestrian dataset and the extended KITTI dataset show that ECA-Res2Net-YOLO is higher than other pedestrian detection algorithms in terms of recall rate and average precision,the recall rate is87.6%and 69.6%,respectively,and the average precision is 85.5%and 70.9%,respectively.At the same time,compared with the original YOLOv3,the computational complexity of the model and the amount of model parameters are significantly reduced,and the detection speed reaches 27 FPS.While maintaining the detection accuracy,it also meets the requirements of the real-time model.（2）In view of the problems that multiple downsampling and convolution may lead to the loss of small target feature information and the insufficient feature fusion of YOLOv3 feature fusion network,an improved YOLOv3 feature fusion network is proposed.Firstly,an improved spatial pyramid pooling structure DSPP module is proposed,which draws on the idea of spatial pyramid pooling and dense connection,and integrates the extracted features by adding 1×1convolution after the maximum pooling layer.At the same time,the input of the pooling layer with large pooling kernel contains not only the original input,but also the output of the pooling layer with small pooling kernel after 1×1 convolution integration.The improved spatial pyramid pooling structure DSPP module enhances the expressive ability of features and effectively solves the problem of feature loss in deep networks.Secondly,the idea of Bi FPN is used to improve the YOLOv3 feature fusion network.On the basis of the FPN structure,a bidirectional feature fusion channel is constructed to fully fuse the extracted shallow and deep features.At the same time,for the F₄ layer features that are used many times,a horizontal skip connection is added in the same layer of the F₄ layer features.So that the network can reuse the extracted F₄ layer features,and the obtained feature map semantics are more sufficient.Finally,experiments are carried out on the used experimental dataset and compared with other pedestrian detection algorithms.The experimental results show that the YOLOv3 model of the improved feature fusion network can effectively improve the accuracy of pedestrian detection.