Accurately counting vehicle turns at busy intersections is a cornerstone of modern urban planning and traffic signal optimization. However, automating this task using single-camera feeds remains a significant challenge due to occlusions, adverse weather, and perspective distortions. This paper, presented at CVPR 2020 workshops, introduces an integrated pipeline that combines deep learning-based detection with traditional background modeling and a sophisticated trajectory matching schema to achieve high accuracy and real-time performance.
The Challenge of Intersection Monitoring
Standard computer vision approaches often struggle in real-world traffic scenarios. Several factors contribute to this difficulty:
- Occlusions: Vehicles frequently obscure one another, especially in heavy traffic.
- Environmental Factors: Heavy rain, glare, and shifting shadows can confuse even advanced neural network detectors.
- Perspective Distortion: High-angle "floating" camera views make it difficult to distinguish specific lanes.
- Fragmented Tracks: Temporary loss of a vehicle by a tracker leads to "broken" trajectories, causing double-counting or missed counts.
An Integrated Pipeline for Turn-Counting
The authors propose a sequential pipeline consisting of five integrated modules designed to move beyond "one-size-fits-all" models.
1. Hybrid Vehicle Perception
To improve recall in extreme weather, the system combines a high-performance deep learning detector (NAS-FPN) with GMM-like background modeling. While deep learning is the industry standard, the authors found that traditional background subtraction provides a necessary "safety net," capturing moving objects that a neural network might miss due to low visibility or unusual lighting.
2. Robust Tracking and Trajectory Modeling
The system utilizes DeepSort for multi-object tracking, enhanced by a custom post-processing step to smooth trajectories and filter out static noise like parked cars. A key innovation is the creation of a lane-level gallery. By aggregating a large number of "clean" trajectories, the system builds a template of valid movements for a specific intersection.
3. Spatial Trajectory Matching
Most trackers rely on timing, but traffic lights make vehicle duration unpredictable. Instead, the authors segment paths based on spatial distance. This allows the system to match a noisy, real-time path to the most likely template in the lane-level gallery, effectively "snapping" fragmented tracks to valid movements.
4. Fragment Stitching
To address broken tracks, the system calculates a vehicle's "survival time." If two short trajectory fragments have matching survival times and non-conflicting velocities, they are merged and treated as a single vehicle.
Performance and Results
The method was evaluated on the AICity 2020 Track1 dataset, which includes 20 real-world scenes across 31 video sequences featuring various weather conditions and complex traffic.
Tested on a Tesla P40 GPU, the system achieved:
- S1-Score: 92.60% (combining detection and counting accuracy).
- Effectiveness: 91.40%.
- Speed: 13 FPS, making it viable for near-real-time monitoring.
Conclusion and Future Work
The integrated approach demonstrates that combining modern deep learning with traditional spatial geometry and background modeling creates a robust solution for traffic analysis. While highly effective, the authors note that identity switches still occur during complex, long-duration occlusions (e.g., a large truck blocking a car). Future work may involve incorporating visual appearance features to better re-identify vehicles after they emerge from obstructions.
Summary updated