Research On Anchor-free Pedestrian Detection Method Based On Multiple Keypoints

In recent years,with the vigorous development of China’s new energy automobile manufacturing industry,autonomous driving technology has shown a completely practical trend,and its technical core is a real-time road pedestrian detection system with high precision and low delay.However,there are usually many complex factors in real road scenes,such as objects blocking each other and long-distance small-scale targets.These complex factors will lead to inaccurate detector positioning or false detection,which will lead to a sharp decline in the performance of pedestrian detection system.In order to solve the above problems,this paper focuses on the detection of serious occlusion and small-scale targets in complex scenes from the perspectives of multiple keypoints detection results fusion and detection process fusion,so as to achieve the purpose of accurately and quickly detecting pedestrians.The main work is summarized as follows:(1)To address the problems that classical pedestrian detection algorithms have high computational complexity and cannot be applied to real traffic road scenarios in real time,this paper selects Center Net based on the anchor free target detection framework as the benchmark algorithm to build an anchor free pedestrian detection system.The system is an end-to-end target detection system,and DLA34 is chosen as the backbone feature network to extract target feature information.The system refers to the Center Net detection head structure and decodes the detection frames and confidence levels of the pedestrian targets in turn.The detection head relies directly on keypoints to predict detection results and suppresses interfering targets through peak point extraction,thus eliminating the need for other cumbersome post-processing operations.(2)To address the problem of severe occlusion and small-scale targets on real roads,which lead to missed and false detection by detectors,this paper proposes a pedestrian detection method based on the combination of dual keypoints.The method firstly introduces deformable convolution on the deep aggregated backbone feature network to expand the perceptual field and enhance the semantic information of human patterns;then,keypoints in the head and centre regions of the human body are used to enhance the semantic information of severely occluded and small-scale targets respectively,so as to effectively extract and fuse the discriminative semantic features of pedestrians,thus significantly reducing the missed detection rate of pedestrians.Finally,the redundant detection results are removed by the merging algorithm.(3)This paper proposes an adaptive method for selecting the optimal keypoints for complex scenes in which the detector is unable to accurately locate the keypoints of the target due to the variable pose of the pedestrian.By expanding the number of target keypoints and adaptively selecting the optimal keypoints for different complex scenarios,the method improves the robustness of the model in dealing with severe occlusions in complex scenes.Due to the increase in the number of keypoints,the DIOU loss function is also introduced to improve the matching effect between keypoints in order to avoid adjacent prediction frames from affecting each other and further enhance the detection accuracy of the algorithm.This keypoint selection mechanism also eliminates the need for additional post-processing algorithms,thus maintaining a high inference speed.In summary,this paper investigates pedestrian detection algorithms based on single,double and multiple keypoints respectively,and conducts comparison experiments with other advanced algorithms on three publicly available pedestrian datasets,Cityperson,Crowd Human and Widerperson.The experimental results show that both result and process fusion using human centre and head keypoints can better balance detection accuracy and speed,which can meet the practical application requirements and have good application in the field of autonomous driving.