Research On Multi-keypoint Anchor-free Pedestrian Detection Algorithm In Complex Scenes

The task of pedestrian detection plays an important role in computer vision,not only as a separate application to achieve value,but also as a foundation to serve other applications,such as pose estimation,behavior recognition,etc.In this study,a multicenter point prediction branch feature extraction model will be designed for occlusion scenarios to avoid missed detections caused by occlusion,and supplemented by a channel non-maximum suppression strategy to remove redundant detection frames caused by multi-branch prediction.Dimensional pedestrian design triggers attention mechanism to reinforce pedestrian regions on feature maps.For occlusion scenarios,this thesis proposes a new bounding box encoding method to solve the problem of missed detection caused by occlusion.In this thesis,the pedestrian area coding is divided into three Gaussian center points,and multi-center point prediction branches are set to predict the location of pedestrians,which are the upper half center point prediction branch,the middle center point prediction branch and the lower half center point prediction branch.These three branches can effectively identify pedestrians in various occlusion situations.In the CityPersons dataset,under the premise of ensuring the performance of the Reasonable subset,the missed detection rate(MR-2)of the heavily occluded Heavy subset is reduced from 49.63%to 45.57%.Due to the setting of multiple center point prediction branches in parallel,redundant detection boxes are generated.This thesis proposes a de-redundancy method to dynamically adjust the non-maximum suppression(NMS)threshold based on channel sources.Due to the multi-branch setting,detection frames that overlap but come from different branch sources are more likely to represent the same pedestrian,so appropriately lowering the NMS threshold between detection frames from different branch sources will effectively filter redundant detection frames.For small-size pedestrian detection,this thesis proposes a trigger attention mechanism.In addition to the three area center point prediction branches,a visible area center point prediction branch is generated in parallel,and the weight factor is established by the similarity between the feature map of the visible area center point prediction branch and the three area center point prediction branch feature maps.When the similarity between the feature map predicted by a region and the visible region is higher,it means that the feature map of the region is more credible,and the feature map of the region is enhanced according to the weight factor.With the enhancement of the trigger attention mechanism,the model is reduced by 0.94%in the Heavy subset MR-2.