Research On Pedestrian Detection Based On Anchor-free Detection Algorithm

The pedestrian detection problem has been one of the hot spots of research in the field of object detection for many years due to its rich application scenarios in real life,especially in the field of intelligent driving.However,pedestrians in real-life scenarios are often diverse in scale and often appear to obscure each other,increasing the difficulty of performing effective detection.With the development of deep learning,most of the methods that work well in the field of pedestrian detection are two-stage anchor-based detection models,which rely on the preset anchors to generate proposal regions and can achieve high detection accuracy.However,these models need to adjust the hyperparameters related to anchors,and the network structure is more complex,which affects the detection speed.While the one-stage detection model is faster,but the detection accuracy is not high enough to meet the demand on accuracy.In the past few years,a series of anchor-free object detection models have been proposed to ensure detection speed while obtaining detection accuracy comparable to or even better than that of the two-stage detection models,but there are still relatively few related studies of anchor-free detection model in the field of pedestrian detection.To address the problem that accuracy of current anchor-free detection model is not high enough for pedestrian detection task,especially for small-scale pedestrian detection,this thesis proposes an anchor-free pedestrian detection model.The proposed model use a multi-scale feature fusion module based on attention mechanism,by upsampling and concate the output features of different scale output layers of the backbone network,and then use a spatial channel attention module to weight the fused features,which improves the feature-extracting ability of the model for multi-scale features.The final^-2metrics of the anchor-free pedestrian detection model reach 10.9%and 13.7%on the Reasonable subset of City Persons dataset and the Small subset of small-scale pedestrians,respectively.To address the problem that the current anchor-free pedestrian detection model is prone to false detection and missing detection for occluded pedestrians in crowded scenes,this thesis designs an anchor-free pedestrian detection model for crowded scenes.In the proposed model,the backbone network is replaced with a deep layer aggregation network with stronger feature extraction capability,and a deformable convolution is introduced to enhance the modeling capability of deformed objects.In addition,a density prediction branch is added in the detection head to predict the maximum intersection over union at each location of the image,thus corresponding to the pedestrian density at each location.Further,a density-based non-maximum suppression algorithm is proposed to use the density information to set the threshold for NMS adaptively,thus improving the missing detection in crowd scenes.The final obtained anchor-free detection model for crowded scenes achieves^-2metrics of 9.1%and 46.7%on the Reasonable subset of the City Persons dataset and the heavily occluded Heavy subset,respectively.