Scholarlink

<p>The paper describes the current state of development in multi-object tracking and comprehensively reviews the most recent algorithmic achievements, applications, challenges, and directions for future research. This paper discusses different MOT methods across various domains, including autonomous driving, surveillance, robotics, and intelligent transportation systems, and how the paradigm switches from old tracking methods to deep learning-based approaches. The development of the MOT models based on CNNs, RNNs, graph-based tracking, and transformer architectures is reviewed. This work critically evaluates the problems related to MOT, including occlusions, MOTion blur, identity preservation, and real-time performance constraints for demanding road applications in resource-scarce environments. Additionally, 152 MOT models and frameworks’ performances against several widely acknowledged benchmark datasets are evaluated, including MOTChallenge, KITTI, Waymo Open Dataset, and nuScenes under standard metrics, Multiple Object Tracking Accuracy (MOTA), Multiple Object Tracking Precision (MOTP), Identity Switches (IDSW), and Fragments (FRAG). Finally, emerging research directions, such as multimodal sensor fusion in LiDAR and radar, self-supervised learning, domain adaptation, as well as the use of edge computing to conduct real-time tracking, are also discussed. Hybrid tracking strategies to bridge the gap between Deep Learning and probabilistic MOTion models are also discussed, and obtain better robustness but poorer scalability. This review introduces the MOT research bar and poses strategic directions to scalable, efficient, and interpretable tracking solutions for intelligent perception systems for future innovations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Surveillance to self-driving: a comprehensive review of object detection and tracking paradigms

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Puja Das,
  • Chitra Jain,
  • Kamal Kumar Gola

摘要

The paper describes the current state of development in multi-object tracking and comprehensively reviews the most recent algorithmic achievements, applications, challenges, and directions for future research. This paper discusses different MOT methods across various domains, including autonomous driving, surveillance, robotics, and intelligent transportation systems, and how the paradigm switches from old tracking methods to deep learning-based approaches. The development of the MOT models based on CNNs, RNNs, graph-based tracking, and transformer architectures is reviewed. This work critically evaluates the problems related to MOT, including occlusions, MOTion blur, identity preservation, and real-time performance constraints for demanding road applications in resource-scarce environments. Additionally, 152 MOT models and frameworks’ performances against several widely acknowledged benchmark datasets are evaluated, including MOTChallenge, KITTI, Waymo Open Dataset, and nuScenes under standard metrics, Multiple Object Tracking Accuracy (MOTA), Multiple Object Tracking Precision (MOTP), Identity Switches (IDSW), and Fragments (FRAG). Finally, emerging research directions, such as multimodal sensor fusion in LiDAR and radar, self-supervised learning, domain adaptation, as well as the use of edge computing to conduct real-time tracking, are also discussed. Hybrid tracking strategies to bridge the gap between Deep Learning and probabilistic MOTion models are also discussed, and obtain better robustness but poorer scalability. This review introduces the MOT research bar and poses strategic directions to scalable, efficient, and interpretable tracking solutions for intelligent perception systems for future innovations.