Abstract
Detecting walking persons in video sequences is often complicated by complex backgrounds, illumination changes, and shadows. This paper introduces a new adaptive detection algorithm that integrates Gaussian Mixture Models (GMM), Canny edge detection, and continuous frame difference. By operating in both temporal and spatial domains, the algorithm generates a complete initial contour of moving targets, effectively addressing common issues like inaccurate detection, incomplete edges, cavitation, and ghosting. Experimental results demonstrate improved robustness and real-time performance.
Introduction
Detecting a moving target, such as a walking person, is a fundamental task in computer vision, yet it remains challenging due to environmental noise. Traditional methods often suffer from specific drawbacks:
- Optical Flow is accurate but computationally heavy.
- Inter-frame Difference is fast but prone to "cavitation," where the center of a moving object is missed because pixel values remain static between frames.
- Background Subtraction fails under dynamic lighting or busy backgrounds.
This research proposes a hybrid approach designed for "big data" video streams that combines these techniques to mitigate their individual weaknesses.
Methodology
The proposed algorithm processes video data through a multi-path pipeline that fuses temporal and spatial information to create a robust target mask.
Temporal and Spatial Modeling
- Gaussian Mixture Model (GMM): Used for temporal modeling to maintain and update the background dynamically, allowing the system to adapt to lighting changes.
- Canny Edge Detection with Otsu’s Method: For spatial modeling, the Canny operator is enhanced with Otsu’s adaptive thresholding. This allows the system to calculate optimal edge thresholds based on real-time image contrast.
- 7-Frame Continuous Difference: Instead of simple two-frame subtraction, a 7-frame sequence is used. "AND" operations across these frames filter out transient noise, ensuring only consistent motion is captured.
Fusion and Post-Processing
The results from the GMM-based contour and the combined frame-difference/edge detection contour are merged using a logical "OR" operation. This fusion ensures that the hollow centers (cavitation) missed by frame difference are filled by the GMM, while the precise edges missed by GMM are provided by the edge detection path. Finally, RGB color space analysis is used for shadow suppression, followed by morphological operations (dilation and erosion) to produce the final target silhouette.
Key Results
The hybrid algorithm was tested across various environments, including corridors, store entrances, and gyms.
- Accuracy: In corridor tests, the algorithm achieved a 91.3% correct detection rate, significantly higher than the 76.1% achieved by standard three-frame difference methods.
- Visual Integrity: Qualitative analysis confirmed that the "OR" fusion successfully eliminated cavitation and ghosting, producing solid silhouettes.
- Performance: While more complex, the algorithm maintained a processing speed of 0.57 to 0.63 seconds per frame in MATLAB, proving its viability for near real-time applications.
Conclusion
The mixed adaptive detection algorithm provides a robust solution for moving target detection in complex surveillance scenarios. By leveraging the complementary strengths of GMM and frame difference, it achieves high accuracy and visual integrity. Future work will focus on further optimizing the computational efficiency to handle high-resolution "big data" streams more effectively.
Summary updated