Multi-Target Tracker for Low Light Vision
Nadya Abdel Madjid, Arjun Sharma, Bilal Hassan, Naoufel Werghi, Jorge M.M. Dias, Majid Khonji · 2023
Recently, remarkable progress has been achieved in addressing the problem of multi-object tracking (MOT), especially in the context of autonomous vehicles (AV). One of the prospective domains of MOT tracking is thermal infrared (TIR) tracking, which can equip an AV with the ability to track pedestrians and vehicles in low light conditions. In this paper, we propose a multi-object tracker for TIR images with a focus on simple and light-weight algorithmic solution. We base our solution on DeepSORT algorithm and extend it to TIR tracking of both pedestrians and vehicles. To adopt DeepSORT algorithm, we design an appearance descriptor suitable for the association problem for TIR images. Furthermore, to address the problem of missing association and detection, we propose a fusion block to merge short tracklets belonging to the same object in one track. We evaluate the tracker on CAMEL dataset and experimentally on the sequences we collected using an IR-camera. The tracker's code is available at github.com/AV-Lab/IR_tracking.