Multi-Object Tracking

MOT Benchmark is one of the most popular benchmarks for object tracking. It consists of a diversified set of videos for training and testing. The goal is to give to each object the same correct ID during the video. It is a difficult task for an algorithm to recognize the same person in a crowded environment with a lot of occlusions and appearance changes. In this benchmark one can find the state-of-the-art methods applied and competing for the best accuracy.

I developed a method that precisely localizes objects and predicts IDs. At the time of my testing (mid July 2020), my method performed number one under the category of supervised, published and online tracker using public detections. Results and videos can be found here.

I also include a demo video here: