Abstract
Visual object tracking remains a challenging task due to abrupt motion, occlusions, and significant appearance deformations. PACNet addresses these issues by introducing a hierarchical deep reinforcement learning (DRL) framework that treats tracking as a multi-agent decision-making process. By combining a high-level policy agent for strategic mode switching with a low-level actor-critic agent for precise bounding box regression, PACNet achieves state-of-the-art performance while maintaining real-time processing speeds of 40 FPS.
Introduction
The goal of visual object tracking is to localize a target in a video sequence based only on its initial position. Traditional deep learning trackers typically fall into two categories: classification/Siamese-based methods and standard RL-based methods. Siamese trackers often rely on inefficient search strategies like random sampling or cross-correlation, which can ignore temporal context. Conversely, standard RL trackers frequently use discrete actions (e.g., move left, move right), requiring numerous steps to converge, or single-step continuous actions that struggle with complex motion.
PACNet aims to bridge the gap between machine tracking and human-like visual attention. Humans track by using strategic search steps and knowing when to reset or update their focus. PACNet models this behavior as a hierarchical Markov Decision Process (MDP).
Methodology: The PACNet Framework
PACNet consists of three core components designed for collaborative tracking:
- Siamese Observation Network: Utilizing a shared ResNet-18 backbone, this network extracts features from both previous and current frames. An integrated LSTM layer captures temporal relationships, enhancing the tracker's robustness against appearance changes and occlusions.
- High-Level Policy Agent: This agent manages the tracking strategy by making discrete decisions about the tracking status. It selects from four actions:
search(continue tracking),stop(target localized),update(refine the model), orreinit(restart tracking). - Low-Level Actor-Critic Agent: When the Policy Agent initiates a
search, the Actor-Critic agent performs continuous bounding box regression (Δx, Δy, Δw, Δh). This continuous action space allows for precise localization in significantly fewer steps than discrete-action models.
Expert Guidance and Temporal Memory
A key innovation in PACNet is its use of "expert guidance" for update and reinit actions. When the agent is uncertain, it leverages a high-performance "expert" tracker (such as SiamRPN++ or DiMP) to adjust the bounding box. Furthermore, the inclusion of an LSTM allows the network to maintain a memory of the target's motion history, which is critical when the target's immediate appearance is misleading due to occlusion.
Experimental Results
PACNet was trained on the GOT-10k dataset and evaluated across five major benchmarks: OTB-100, UAV-123, LaSOT, VOT-2019, and the GOT-10k test set.
Performance and Efficiency
The hierarchical approach allows PACNet to maintain a high frame rate of 40 FPS on a single GTX 2080 Ti, meeting real-time requirements without sacrificing accuracy.
- OTB-100: Achieved a precision of 0.876 and an AUC of 0.670, outperforming trackers like ACT and ECO-HC.
- UAV-123: Reached an AUC of 0.620, surpassing SiamRPN++ (0.615) and DaSiamRPN (0.569).
- LaSOT: Ranked second with a 0.553 AUC, proving more efficient than traditional deep trackers like DiMP-50.
- VOT-2019: Achieved an Expected Average Overlap (EAO) of 0.300, demonstrating high robustness.
Conclusion and Future Work
PACNet demonstrates that hierarchical collaboration between agents can significantly improve tracking efficiency and accuracy. By using continuous action spaces for regression, the tracker locates targets in just 1–3 steps, a major improvement over previous RL-based trackers.
While highly effective, PACNet's peak performance is currently influenced by the quality of the expert model used during the update phase. Future research may explore meta-reinforcement learning to enable faster parameter adaptation during online inference, potentially improving performance on entirely unseen object classes.
Summary updated