Algorithmic and Challenge-Based Deep Learning Methods for Real-Time Multi-Object Tracking

Abstract

Multi-object tracking (MOT) is the challenge of tracking the status of an unknown and time-varying number of objects using noisy observations. Among other fields, this topic has important applications in autonomous driving, defensive systems, and tracking animal activity. Whether the MOT job is model-based or model-free depends on the availability of accurate and controllable models of the environment. Model-based MOT's Bayes-optimal closed-form solutions enable it to achieve SOTA performance. The performance of these techniques is limited since their feasibility requires approximation in challenging scenarios. Deep learning (DL) techniques offer a good alternative, however existing DL models are almost exclusively designed for model-free contexts and are challenging to adapt to model-based situations. First, the definition of multi-object tracking, its background, and the application benefits of deep learning and different stages of image tracking detector are discussed. This is followed by a detailed analysis of the different methods for multi-object tracking. The commonly used assessment metrics and datasets are also presented, after which the experimental results of different methodologies with the proposed tracking detector on the datasets are compared. Finally, the strengths and weaknesses of multi-object tracking methods are analysed and the direction of next study is proposed. Both the development of multi-object tracking and the expansion of social security depend on it.