Improvement of Soft-YOLOv5 Algorithm for Stray Cats and Dogs Detection

Real-time monitoring and instantaneous tracking of stray cats and dogs within a specific area using image recognition and object detection in computer vision technology is currently the predominant trend. Cat and dog recognition, as a classic object detection project, serves the purpose of identifying the quantity and categories of cats and dogs in a single image or video frame. This study builds upon the YOLOv5 framework and addresses the practical challenges of cat and dog detection, such as overlapping and occlusion. To enhance detection accuracy and stability, the Soft-NMS and Merge-NMS algorithms are contrasted with YOLOv5's NMS algorithm. The improved Soft-NMS and Merge-NMS algorithms are evaluated through performance testing and compared with the original NMS algorithm. As a result of these two algorithm enhancements for YOLOv5's NMS algorithm, the Soft-NMS algorithm demonstrates a 2.2% performance improvement over YOLOv5 using the NMS algorithm. Therefore, Soft-YOLOv5 is better suited for the task of detecting stray cats and dogs in urban communities and streets.