a project on multi object tracking
We demonstrated that the bounding box regressor of a trained Faster-RCNN detector is sufficient to address the majority of tracking cases seen in current benchmarks. A detector converted as Tracktor requires no special training in tracking ground truth data and may function online. Furthermore, we demonstrated that our Tracktor can be extended with re-identification (reID) and camera motion compensation (CMC), resulting in a significant new state-of-the-art on the MOT Challenge. We examined the performance of many dedicated tracking approaches on hard tracking circumstances, and none performed any better than our regression-based Tracktor. We hope that our study develops a new tracking paradigm that makes full use of the object detector’s capabilities.
Multiple Object Tracking (MOT) plays a vital role in the technical area of computer vision. MOT is quite a fascinating problem with a varied range of applications, especially concentrating on security aspects. It deals with the process of positioning and tracking object instances over time using a video as the input. The task of MOT includes selecting objects of interest, locat�ing multiple objects, maintaining their identities, the association of objects, and motion prediction.
This project is run using Google Colab.