MinkowRadon: Multi-Object Tracking Using Radon Transformation and Minkowski Distance

The latest trend in multiple object tracking (MOT) is bending to utilize deep learning to improve tracking performance. With all advanced models such as R-CNN, YOLO, SSD, and RetinaNet, there will always be a time-accuracy trade-off which puts constraints to computer vision advancement. However, it is not trivial to solve those kinds of challenges using end-to-end deep learning models, adopting new strategies to enhance the aforementioned models are appreciated. In this paper we introduce a novel radon transformation based framework, which takes advantage of color space conversion and squeezes the MOT problem to signal domain using radon transformation. Afterwards, the inference of Minkowski distance between sequence of signals is used to estimate the objects' location. Adaptive Region of Interest (ROI) and thresholding criteria have been adopted to ensure the stability of the tracker. We experimentally demonstrated that the proposed method achieved a significant performance improvement in both The Multiple Object Tracking Accuracy (MOTA) and ID F1 (IDF1) with respect to previous state-of-the-art using two public benchmarks.