Abstract
Real-time and accurate vehicle tracking is a fundamental requirement for Intelligent Transportation Systems (ITS), providing the data necessary for traffic condition evaluation and command dispatch. This paper proposes a Multi-Feature Fusion Particle Filter (MFFPF) designed to overcome common tracking challenges such as occlusion, illumination changes, and similar target interference. The method fuses improved color local entropy (CLE) with Scale-Invariant Feature Transform (SIFT) features, utilizing vehicle symmetry to enhance robustness. A Kalman filter is integrated to predict motion states and assist in target re-acquisition. Experimental results demonstrate that the proposed method significantly improves tracking accuracy and real-time performance in complex traffic environments.
Introduction: Challenges in Traffic Surveillance
Reliable vehicle tracking in surveillance video is often hampered by environmental complexities. Standard trackers frequently lose targets when they move through shadows, are partially hidden by other vehicles, or are surrounded by visually similar cars.
The authors identify three primary hurdles:
- Illumination Variability: Sudden changes in lighting (e.g., shadows) can cause feature drift.
- Target Occlusion: Large vehicles or infrastructure can temporarily hide the target.
- Real-Time Constraints: Complex feature extraction must be optimized for live surveillance applications.
Methodology: Multi-Feature Fusion and Symmetry
The proposed MFFPF algorithm addresses these hurdles by combining complementary features and geometric priors.
Improved Color Local Entropy (CLE)
Traditional color histograms lack spatial context. By incorporating local entropy, the algorithm captures the texture and spatial distribution of pixels within the target region. To maintain real-time performance, the authors utilize a Taylor series expansion to simplify logarithmic calculations. Furthermore, the inherent symmetry of vehicles is exploited; if one side of the vehicle is occluded, the system relies on the visible side's entropy to maintain the model.
Optimized SIFT Features
While SIFT is robust to scale and rotation, it is computationally intensive. The authors optimize SIFT by weighting feature points based on their temporal persistence and using symmetry to replicate eigenvectors, which reduces computation while increasing reliability.
State Prediction and Target Recovery
A Kalman filter is employed to predict the vehicle's motion state, providing a search window even when the target is temporarily lost. If re-acquisition is necessary, the system combines the predicted position with license plate recognition and SIFT matching to quickly locate the correct vehicle.
Experimental Results
The MFFPF algorithm was evaluated using the Object Tracking Benchmark (OTB), the MIT Traffic Dataset, and custom surveillance sequences.
- Benchmark Performance: On the OTB dataset, the MFFPF achieved second place in both average precision and success rate among ten state-of-the-art trackers, including KCF and CCOT.
- Robustness to Occlusion: Qualitative tests showed the system could maintain or rapidly resume tracking even after total occlusion, thanks to the fusion of motion prediction and unique feature matching.
- Real-Time Efficiency: The optimizations in CLE and SIFT calculations allowed the tracker to meet the requirements for real-time surveillance applications.
Conclusion
By fusing color local entropy and SIFT features with geometric symmetry and predictive filtering, the MFFPF provides a robust solution for vehicle tracking in challenging traffic scenarios. Future work will focus on improving performance in low-light nighttime conditions where feature extraction becomes significantly more difficult.
Summary updated