Abstract
Moving target detection plays an important role in computer vision. However, traditional algorithms such as frame difference and optical flow usually suffer from low accuracy or heavy computation. Recent algorithms such as deep learning-based convolutional neural networks have achieved high accuracy and real-time performance, but they usually need to know the classes of targets in advance, which limits the practical applications. Therefore, we proposed a model free moving target detection algorithm. This algorithm extracts the moving area through the difference of image features. Then, the color and location probability map of the moving area will be calculated through maximum a posteriori probability. And the target probability map can be obtained through the dot multiply between the two maps. Finally, the optimal moving target area can be solved by stochastic gradient descent on the target probability map. Results show that the proposed algorithm achieves the highest accuracy compared with state-of-the-art algorithms, without needing to know the classes of targets. Furthermore, as the existing datasets are not suitable for moving target detection, we proposed a method for producing evaluation dataset. Besides, we also proved the proposed algorithm can be used to assist target tracking.
Introduction
In computer vision, detecting moving targets is a foundational task for applications ranging from autonomous driving to security monitoring. Current techniques face a significant trade-off. Traditional methods like frame differencing or optical flow are computationally efficient but sensitive to camera shake and lighting changes. Conversely, modern deep learning-based detectors are highly accurate but typically class-specific; they may fail to detect moving objects that do not belong to their pre-trained categories (e.g., debris or unknown animals).
To bridge this gap, this research introduces a class-agnostic approach. By focusing on the act of motion and consistency of appearance rather than semantic labels, the system can detect any moving object regardless of its category.
Methodology
The proposed detection pipeline is divided into three distinct stages, leveraging deep learning features and probabilistic modeling.
1. Extracting the Moving Area
Instead of comparing raw pixels, which are prone to noise, the algorithm processes adjacent video frames through a pre-trained VGG-16 network. By calculating the difference between high-level feature maps (specifically layer 14), the system identifies structural changes. This approach is significantly more robust to camera vibrations and illumination shifts than traditional pixel-level differencing.
2. Predicting Color and Location
Once a rough moving area is identified, the system employs Bayesian estimation to build two probability maps:
- Color Probability Map: This map determines the likelihood that a pixel belongs to the target based on the color distribution within the moving area versus the background.
- Location Probability Map: A Gaussian distribution is generated, centered on the initial moving area, to provide spatial constraints.
3. Target Detection via SGD
The final "target probability map" is created by multiplying the color and location maps. To determine the optimal bounding box, the researchers use Stochastic Gradient Descent (SGD). This iterative process refines the box’s position and size to maximize the target score while penalizing excessively small boxes, allowing for precise localization.
Evaluation and Results
Dataset Production
Because existing datasets were either too granular (pixel-level) or class-restricted, the authors developed a new evaluation dataset derived from the OTB100 and VOT2017 benchmarks. This dataset includes over 1,500 image pairs from 110 sequences, simulating various frame rates and focusing on single-target detection.
Performance Metrics
The proposed method demonstrates superior performance and efficiency:
- Accuracy: Achieved a Success Rate (AUC) of 0.457, outperforming Faster R-CNN (0.307) and Flow-net based methods (0.227).
- Speed: The system operates at 40 frames per second (fps) on standard hardware (i7 CPU, 1060Ti GPU), meeting real-time requirements.
- Tracking Utility: When integrated with existing trackers like KCF or CCOT, it improved precision by up to 0.61%, particularly in scenarios involving occlusion.
Conclusion
This work presents a robust, real-time solution for class-agnostic moving target detection. By utilizing CNN features for motion detection and Bayesian refinement for appearance modeling, the algorithm avoids the limitations of class-specific detectors. While currently optimized for single targets, future work will explore multi-target detection and end-to-end integration of the pipeline to further enhance performance in complex environments.
Summary updated