Adaptive Object Tracking via Multi-Angle Analysis Collaboration

Wanli Xue, Zhiyong Feng, Chao Xu, Zhaopeng Meng, Chengwei ZhangSensors 2018 · OpenReview

TL;DR

This paper introduces MACT, a novel object tracking framework that mimics human visual tracking by analyzing the scene from multiple perspectives. By employing reinforcement learning to dynamically select features and predict object motion, MACT achieves improved accuracy and robustness compared to traditional methods.

What Problem Does This Paper Address?

Visual tracking, the task of locating an object in a video sequence, faces numerous challenges like illumination changes, occlusions, and fast motion. While current methods often rely on complex deep learning models or fixed feature sets, this paper argues for a return to the essence of tracking: simulating human visual perception. Humans don't just process raw pixels; they actively focus attention and anticipate movement. Existing trackers often lack this dynamic, multi-faceted approach, leading to performance limitations in complex scenarios.

Methods

The core of MACT is a "multi-angle analysis collaboration tracking" (MACT) framework that combines a basic tracking model with a strategic framework.

  • Basic Tracking Framework: This component, inspired by existing work, includes standard elements like a motion model, feature extractor, observation model, model updater, and ensemble post-processor. However, it's modified to support the strategic framework.
  • Strategic Framework: This is the novel contribution, implemented using multi-dimensional state-action space reinforcement learning (specifically Q-learning with an ε-greedy exploration strategy). It models two "thinking angles":
  • Observer's Attention: This corresponds to selecting the most appropriate image features (raw grayscale, raw color, or HOG) for the current scene.
  • Object's Motion: This corresponds to predicting the object's movement trend, guiding the motion model for more purposeful sampling.

The reinforcement learning agent learns strategies to optimize tracking quality by receiving rewards based on tracking success, feature selection, and movement prediction.

Datasets

The primary dataset used for evaluation is the OTB50 dataset. This dataset is described as a small but comprehensive collection of 50 annotated video sequences that cover a wide range of tracking challenges.

Key Insights and Ideas

  • Human-like Tracking: The paper strongly advocates for simulating human visual tracking, emphasizing that humans don't rely on a single, fixed approach but adapt their focus and predictions.
  • Multi-Angle Analysis: The core idea is that analyzing tracking from multiple perspectives (observer's attention and object's motion) leads to more robust and accurate tracking.
  • Reinforcement Learning for Strategy: Using RL to learn dynamic strategies for feature selection and motion prediction is a key enabler for the multi-angle approach.
  • Cooperation is Key: The MACT framework demonstrates that the synergy between a strategic framework and a basic tracking framework significantly enhances performance.

Key Results

  • Accuracy Improvement: MACT showed significant improvements in tracking accuracy, particularly in challenging scenarios. For instance, when compared to the "Basic" tracker, MACT_MTS (with movement trend strategy) outperformed it in OPE success and precision plots across various attributes.
  • Speed: MACT_FSS (with feature selection strategy only) achieved the best tracking speed among the tested methods (21.445 fps), indicating that feature selection can improve efficiency. MACT itself was slightly slower but still competitive.
  • Robustness: MACT demonstrated robustness in challenging video sequences where other methods (like MUSTer) struggled, particularly in cases of fast motion or similar interferences.
  • Comparison to State-of-the-Art: MACT achieved competitive results against mainstream tracking methods, often leading in several attributes on the OTB50 benchmark. It also showed good performance on the VOT-2015 dataset, outperforming DSST and MEEM in accuracy, overlap, and speed.

Limitations and Future Work

  • Limited Angles: The current MACT framework models only two "thinking angles" (observer's attention and object's motion). The authors suggest this framework can be extended to incorporate more perspectives for even better performance.
  • Computational Cost: While MACT aims for efficiency, the use of features like HOG can still be computationally intensive, especially in scenarios requiring complex feature analysis.
  • MUSTer Comparison: In some qualitative comparisons, MACT was noted as being slightly inferior to MUSTer, particularly in sequences with severe occlusion or very similar interferences. However, MACT showed advantages in other challenging sequences due to its movement trend strategy.

Future work could involve expanding the number of "thinking angles" and further optimizing the feature selection and motion prediction mechanisms.

Summary updated

BibTeX
@inproceedings{xue2018adaptive,
  author = {Wanli Xue and Zhiyong Feng and Chao Xu and Zhaopeng Meng and Chengwei Zhang},
  title = {{Adaptive Object Tracking via Multi-Angle Analysis Collaboration}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=sUAs24j_R4},
}
Figure