SimpleTrackV2: Rethinking the Timing Characteristics for Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Detection-based multi-object tracking often relies on Kalman filtering for trajectory prediction.
• Kalman filtering struggles with real-world challenges like camera jitter and temporary target loss.

Purpose of the Study:

• To develop an improved multi-object tracking algorithm, SimpleTrackV2, addressing limitations in state prediction and feature fusion.
• To enhance robustness against camera jitter and target occlusion in complex scenarios.

Main Methods:

• Proposed LSTM-MP for target state prediction, utilizing Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) to encode historical motion.
• Introduced TSA-FF, a spatiotemporal attention-based fusion algorithm, to adaptively enhance target appearance features during occlusion.

Main Results:

• SimpleTrackV2 demonstrated superior performance over the baseline SimpleTrack on the MOT17 dataset.
• Achieved notable improvements in MOTA (1.6%), IDF1 (3.2%), and HOTA (6.1%).
• Ablation studies confirmed the effectiveness of LSTM-MP and TSA-FF components.

Conclusions:

• SimpleTrackV2 offers a more robust solution for multi-object tracking, effectively managing camera jitter and occlusion.
• The proposed LSTM-MP and TSA-FF modules significantly contribute to improved tracking accuracy and reliability.