Convex hull regression strategy for people detection on top-view fisheye images

Due to the severe distortion of the fisheye image, the rectangular bounding box contain a lot of invalid information. So the multi-point representation method are emerging. However, it will fail in some extreme cases especially when the center of the object is not in the instance. In this work, we propose a Convex Hull Regression Strategy for people detection on top-view fisheye images. It replaces the instance with its convex hull to solve the above challenging issue and can be pre-trained on regular datasets without additional processing. In addition, the mosaic and mixup data augmentation methods that perform well under rectangular boxes are applied to our representation. Finally, we improve the label assignment and propose a more reasonable loss function called PDIoU loss so as to focus on the overall IoU between ground truth polygon and predicted polygon. Experimental results demonstrate that our method outperforms state-of-the-art algorithms. Source code is available at https://github.com/xiaoxuebajie/CHRS.