Abstract: To address the issues of false detection and missed detection in foggy weather scenarios where images are blurred and targets are difficult to distinguish, a target detection algorithm that fuses a high-resolution network, named High Resolution Cascade RCNN (HR-Cascade RCNN), is proposed. This algorithm adopts HRNet as the feature extraction network for Cascade RCNN, connects parallel sub-networks with different resolutions to extract multi-scale feature information, thus reducing information loss during downsampling and enhancing the semantic representation of targets. Secondly, the CIoU loss function is used to replace the original Smooth L1 loss function, and a penalty term is introduced to measure the correlation between the aspect ratio of real bounding boxes and detected bounding boxes, thus optimizing the convergence performance of the network, and helping to improve the positioning accuracy of detected bounding boxes. Finally, SoftNMS is adopted to improve the candidate box selection mechanism, reducing the false negative rate in situations such as vehicle occlusion, and enhancing the detection ability of the network. Experimental results on real foggy weather datasets RTTS and synthetic foggy weather datasets Foggy Cityscapes show that compared with the original Cascade RCNN, HR-Cascade RCNN improves mAP by 5.9% and 3% respectively.

Key words: foggy scene, target detection, deep learning, cascade RCNN, high resolution