CAMOT: Content Aware Multi Object Tracking
Ratul Kishore Saha, Rekha Singhal, Manoj Nambiar · 2023
Multi-object tracking in video sequences plays a critical role in various computer vision applications. The primary objective is to accurately localize and track objects across consecutive frames. However, existing approaches often suffer from computational limitations and low frame rates, which hinder real-time performance. In this paper, we propose a content-aware multi object tracking framework that addresses these challenges. We classify videos into slow and fast object content videos depending on the features of objects to be tracked in the frames. We apply the computationally intensive YOLO-DeepSORT algorithm by selectively skipping frames, supplemented by low-cost interpolation between them to achieve tracking performance in all frames. Further, we leverage linear approximate Kalman prediction for slow object content and quadratic interpolation for fast object content videos as low-cost techniques. These innovation leads to 4x (approx) increase in throughput on CPU for Visdrone19 and MOT17 datasets. Our findings provide valuable insights for optimizing object tracking paradigm.