| Summary: | To at the low robustness of the existing model for occluded object detection, an occluded object detection algorithm based on fuzzy sample anchor box IoU Matching degree Deviation Aware(IoU_MDA)is proposed. Firstly, fuzzy samples are defined based on Anchor-based, which reflects the degree of object occlusion. Secondly, IoU_MDA is proposed to quantify the degree of interference experienced by fuzzy samples. Then, IoU_MDA_Loss is constructed based on IoU_MDA, combined with IoU and the balance parameter <inline-formula> <tex-math notation="LaTeX">$\Phi$ </tex-math></inline-formula>. To address class imbalance issues and enhance model generality, intra-class and inter-class fuzzy weights, and fuzzy sample focusing parameters are designed on the basis of the initial IoU_MDA_Loss. An occluded object training scheme is designed based on IoU perception, and non-fuzzy sample weight balancing parameters are constructed. Finally, IoU_MDA_Loss is merged with Focal Loss to obtain IoU_MDA_Focal Loss, simultaneously enhancing the detection performance of fuzzy samples and difficult-to-distinguish samples. The experimental results on the WiderPerson and VOC2007 datasets show that the mAP of IoU_MDA_Loss has increased by 2.04%, 2.36%, and IoU_MDA_Focal Loss has increased by 1.82%, 2.65%, respectively, compared to the baseline model. The detection performance surpasses current mainstream algorithms.
|
|---|