Occlusion Target Detection Based on Improved YOLOv5 Model
Zhe Lin, Dan Chen, Yuanlun Zeng · 2022
To solve the problem that the network is not robust due to the insufficient feature information extraction of occluded objects, an occluded objects detection algorithm combining coordinate attention (CA) mechanism and improved non-maximum suppression (NMS) is proposed in this paper. Based on the YOLOv5 model, the coordinate attention mechanism is first added into the feature extraction layer, which considers both feature information and position information to improve the problem of insufficient feature information extraction due to occlusion. Then, based on Generalized intersection over union (GIoU), the center distance and aspect ratio are considered, Complete-IOU (CIoU) is conducted to replace the original loss function to solve the degradation problem existing in GIoU and the problem that Distance-IoU (DIoU) does not consider aspect ratio. Finally, the traditional NMS only considers Intersection over union (IoU) as an indicator to remove duplicate boxes, which leads to error suppression, kCIoU-NMS is proposed instead of NMS to solve the shortcomings brought by IoU. Tested on a test set with different occlusion rates when the targets are occluded from each other and the background. The accuracy of the model in this paper is above 95% when the occlusion rate is 20% and 40%, and when the occlusion rate is 60%, the detection accuracy is 83.80% and 87.60%, which are 16.00% and 14.40% higher than the original model, respectively. The experimental results show that the method proposed in this paper has good effect and robustness.