Human Tracking in Top-view Fisheye Images with Color Histograms via Deep Learning Detection

Fisheye cameras produce panoramic images. For a while, classical people detection algorithms were not optimal in fisheye images because detection bounding boxes were non-oriented. People detection algorithms for top-view fisheye images have been developed recently. However, these algorithms only detect the people present in the different frames but do not follow them through a video sequence. First, we based our work on the RAPiD (Rotation-Aware People Detection in Over-head Fisheye Images) method to detect people in video frames. Then, in order to track the target throughout the video, we use a comparison method for color histograms based on Bhattacharyya distance. This distance is computed with several histograms with different properties relating to the number of bins or the colorspace to compare the efficiency. Finally, their position is assessed by computing an angle and a distance to the camera. As a result, in a video where several people are detected, we are able to follow the path of one single person throughout the video.