Research On Small Object And Occlusion Object Detection Algorithm Optimized On SSD Model

With the continuous development of image processing technology based on deep learning and the improvement of hardware computing capabilities,object detection technology has made great progress in recent years.The object detection technology can locate and identify the object in the image,and plays a very important role in tasks such as image retrieval,scene understanding and event retrieval.However,in actual application scenarios,there is still room for improvement in the detection performance of the object detection algorithm.Optimizing and improving the detection performance of the object detection algorithm is becoming a hot research and application field in academia and industry,which has important theoretical value and practical significance.In the application scenario of pedestrian detection,the size change of the object and the mutual occlusion between objects are the main factors affecting the detection effect of the algorithm.The size change problem is mainly due to the different distances between the object and the viewfinder,causing different pedestrians to occupy different proportions in the picture;and the occlusion problem is mainly due to the overlap of pedestrians during the movement process,resulting in the loss of some pedestrian characteristics..Aiming at the above problems,this article is based on the SSD(Single Shot multibox Detector)object detection model.By optimizing the model structure and loss function of the SSD,its detection capability for multi-scale objects and scenes with occlusion conditions has been significantly improved.The main research contents and research results of this article are as follows:(1)The technical process of pedestrian detection is summarized,the development process and key technologies from traditional pedestrian detection technology to deep learning-based pedestrian detection technology are summarized,and the current common pedestrian detection framework is emphasized.The characteristics of common detection frameworks such as R-CNN,Fast R-CNN,Faster R-CNN,and YOLO(You Only Look Once)were analyzed and evaluated.(2)The model structure of the SSD algorithm,candidate frame generation and matching strategy,data set processing strategy,and loss function are introduced.By analyzing the multi-scale output structure of the SSD,the reason for the insufficient detection capacity of the SSD model for small objects is determined.This paper proposes a small object feature enhancement module and a 6-scale feature fusion output structure.Experiments are performed on the Pascal VOC dataset and MS COCO data.Experimental results show that the improved SSD object detection model has higher detection accuracy for small objects,and is more robust to different object types and scenarios;(3)An improved occlusion object loss function and C-NMS(Central distance based Non-Maximum Suppression)algorithm based on the Repulsion loss function are proposed.By analyzing the problems of the Repulsion loss function in the scenario where some objects overlap,an improved Repulsion loss function based on the new attraction and exclusion terms is proposed.By analyzing the defects of traditional NMS(Non-Maximum Suppression)algorithm when processing occlusion objects,a C-NMS based on center distance is proposed.And experimentally verified that the SSD algorithm based on improved loss function and C-NMS algorithm has better detection effect in the presence of overlapping detection scenarios.