Research On Anchor-free Pedestrian Detection Algorithm Based On Convolutional Neural Networ

Pedestrian detection technology has always been a hot and difficult point in target detection,and is widely used in intelligent transportation,autonomous driving and intelligent robots.Its task is to collect images of pedestrians in various scene,and mark the location of pedestrians in the images using a positioning box of appropriate size.In this paper,the anchor-free pedestrian detection method based on convolutional neural network is studied,and the scale and occlusion problems in pedestrian detection are discussed in depth.At the same time,considering the real-time performance of the detection algorithm,an improved pedestrian detection algorithm is proposed.This paper improves on the anchor-free object detection algorithm Center Net.Firstly,the DSRes Ne Xt residual structure is proposed,which uses depthwise separable convolution instead of ordinary convolution to reduce the amount of computation;At the same time,the DSCSP module is designed,which can improve the transfer efficiency of gradient information in the network model,optimize the network structure,and further improve the learning ability of the network model.Then,the Multi PNet multi-scale pedestrian detection network is designed to improve the accuracy of the network model to detect multi-scale targets,and use the dilated convolution to increase the receptive field and weaken the influence of target size on the network extraction features,so as to detect multi-scale pedestrian targets more accurately.Finally,a feature processing module is added to the detection head to further improve the detection effect of the network for detecting multi-scale targets.In the training process of the network model,the loss function of the network is optimized to remove the loss of position deviation,which further increases the real-time performance of the network model.Experiments show that the network model designed in this paper can accurately detect multi-scale pedestrian targets while satisfying real-time performance.In order to further optimize the detection effect of the network model on occluded targets,we continue to improve the proposed multi-scale network Multi PNet.First,the GCBAM attention module is designed,which can reduce the network parameters and increase the weight of the network model in the channel and region of interest in the feature map,thereby improving the effect of the network model in detecting occluded pedestrians.Then,according to the characteristics of pedestrian targets,a new feature fusion method FPN-P is designed,which fuses more information in the deep feature maps of the network model to obtain more semantic information,so as to better detect occluded targets.Finally,for the occlusion target,the Grid Mask-P random occlusion algorithm is proposed to simulate the occlusion phenomenon in the image and make the network model pay more attention to the overall characteristics of the target;At the same time,multi-strategy enhancement operations such as optical changes and geometric changes are performed on the image data to increase the robustness of the network model,so that can better detect pedestrians and improve the performance of the network to detect occluded pedestrians.Experiments show that the improvement of the multi-scale pedestrian detection algorithm Multi PNet in this paper can effectively improve the detection performance of the network to detect occluded pedestrians.