Fish-Eye Camera Human Detection Based on YOLO

Object detection is a fundamental task in computer vision. Tremendous attempts has been made to enhance the performance. However, detecting objects with various scales and rotations is still challenging. In this paper, we propose a new method based on a YOLO convolution neural network for human detection in fish-eye cameras. We create different datasets with different scales and rotations and train separate detectors for each scale. Our results show that our approach improves human detection accuracy to 83% while maintaining low false positive rates of nearly 0.