Computationally Efficient Automatic Coast Mode Target Tracking Based on Occlusion Awareness in Infrared Images

So-Hyun Kim, Gwang-Il Jang, Sungho Kim, Junmo Kim

TL;DR

This paper introduces an "occlusion-aware coast mode tracking" (OA-CMT) algorithm designed to improve the robustness of infrared target tracking systems. By intelligently predicting and handling target occlusions, OA-CMT maintains tracking during brief blind periods and efficiently re-acquires the target, outperforming existing methods in computational efficiency and reliability.

What Problem Does This Paper Address?

Tracking targets in infrared (IR) images presents unique challenges. Unlike visible light cameras, IR sensors capture thermal energy, making them effective in low-light conditions and for detecting small objects at a distance. However, IR images often lack distinct features like sharp edges or colors, making traditional feature-based tracking difficult. Centroid trackers, which rely on brightness information, are computationally simple and widely used for IR imagery. However, they are highly susceptible to losing track when targets are obscured by clutter or other objects (occlusion). Existing methods for handling occlusion often require complex setups, are computationally expensive, or degrade performance over time due to "blind updates" of appearance models. This paper aims to address the critical problem of target loss due to occlusion in IR imaging, specifically for applications requiring efficient, real-time tracking on embedded systems.

Methods

The proposed OA-CMT algorithm operates in three main steps:

  1. Occlusion Prediction: This is a two-stage process.
  2. Target Likelihood Obstacle Check: Identifies potential obstacles by searching for background regions with brightness similar to the target. It analyzes "bright blocks" and calculates the "true distance" between the target and potential obstacles.
  3. Background Screening Check: Detects obstacles that might obscure the target even if their brightness differs. This involves analyzing brightness changes in "guard gates" and "background gates" around the target and using a chi-square distribution to determine if a gate is "polluted" by an obstacle. A pre-alarm is triggered by synchronized gate pollution, and a final decision is made based on target gate size reduction.
  4. Memory Tracking (Coast Mode): When occlusion is predicted, the system enters "coast mode." Instead of attempting to track the occluded target, it switches to memory tracking. This uses an α-β-γ filter to predict the target's position based on its historical movement data (location, velocity, acceleration). This filter is chosen for its computational efficiency and suitability for time-invariant states without requiring range information.
  5. Target Re-locking: Once the occlusion is cleared, the algorithm attempts to re-acquire the target. This process uses a target model based on a histogram ratio between the target and background. It searches for candidate blocks, scores them based on similarity to the target model (considering size, intensity, aspect ratio, and probability), and selects the highest-scoring candidate. If the re-locked target is deemed reliable, image tracking resumes.

Datasets

The paper describes using "various test imagery sequences" for computer simulations and real-world testing. These sequences include both infrared (IR) and electro-optical (EO) images, featuring different obstacle types (bushes, street lamps, buildings, humans, traffic signs) and targets. Specific imagery IDs (1-7) are listed in Table 1, detailing the sensor type, obstacle, and description for each sequence. The paper also mentions testing on a naval electro-optical tracking system (EOTS) and an airborne EO/IR system.

Key Insights and Ideas

  • Bridging the Gap for IR Tracking: The core insight is that by combining predictive occlusion detection with a computationally efficient coast mode and a robust re-locking mechanism, the limitations of centroid trackers in IR imagery can be significantly mitigated.
  • Dual-Stage Occlusion Detection: The two-pronged approach (target likelihood and background screening) provides a more comprehensive way to identify potential occlusions, reducing false alarms and improving reliability.
  • Efficient Memory Tracking: The use of an α-β-γ filter for memory tracking is a smart choice for embedded systems, balancing prediction accuracy with low computational overhead during blind periods.
  • Histogram Ratio for Re-locking: Employing a histogram ratio for target re-locking is effective because it leverages intensity information, which is more reliable in IR images than complex feature extraction, and is less sensitive to AGC effects.
  • Computational Efficiency is Key: The paper emphasizes that for real-time embedded applications, computational cost is paramount. The proposed OA-CMT algorithm is designed with this in mind, demonstrating significant speed advantages.

Key Results

  • Computational Efficiency: OA-CMT is shown to be significantly faster than other compared tracking algorithms (TLD, L1-APG, SCM) across various image sizes and target sizes, achieving high frames per second (e.g., 42 FPS for image ID 1, 149.5 FPS for image ID 5).
  • Robustness to Occlusion: Experimental results demonstrate that OA-CMT successfully transitions to memory tracking during occlusions and re-locks the target effectively, while other trackers either lose track, track the wrong object, or exhibit unstable trajectories.
  • Real-World Validation: The algorithm's performance was validated on real naval EOTS and airborne EO/IR systems, confirming its practical applicability.
  • Minimal Abrupt Changes: The predicted and re-locked target positions during occlusion show minimal deviation, which is crucial for smooth servo control in tracking systems.

Limitations and Future Work

  • Limitations:
    • The paper focuses on single-target tracking.
  • While the α-β-γ filter is efficient, it assumes a simplified motion model and does not incorporate range information, which could limit accuracy in certain scenarios.
  • The effectiveness of the re-locking mechanism might depend on the quality of the target model established during normal tracking.
  • Future Work:
    • The authors suggest extending the algorithm to multi-target tracking.
  • Investigating more advanced tracking filters that can incorporate range information could further improve prediction accuracy.
  • Further exploration of adaptive target modeling techniques could enhance re-locking robustness under more challenging appearance changes.

Summary updated

BibTeX
@inproceedings{kim2018computationally,
  author = {So-Hyun Kim and Gwang-Il Jang and Sungho Kim and Junmo Kim},
  title = {{Computationally Efficient Automatic Coast Mode Target Tracking Based on Occlusion Awareness in Infrared Images}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=BHjfu5mP9S},
}
Figure