Abstract
Today, Intelligent Vehicle Detection Systems (IVDS) seek to reduce negative factors such as accidents beyond merely gathering traffic information. This paper proposes a detection algorithm for illegal U-turn vehicles, which can cause critical accidents among road traffic law violations. We hypothesized that if calculated optical flow vectors appeared on an illegal U-turn path, they would likely be caused by illegal U-turn vehicles. To reduce high computational complexity, we utilized the Pyramid Lucas-Kanade algorithm to track specific key-points, such as corners. Furthermore, we detected the center lane using color information and the Progressive Probabilistic Hough Transform (PPHT) to restrict analysis to a specific Region of Interest (ROI). Finally, we selected vectors on the illegal U-turn path and calculated reliability to confirm the detection. Experimental results show that the proposed algorithm is efficient and suitable for real-time applications.
1. Introduction
Illegal U-turns are a major cause of fatal traffic accidents, yet detecting them automatically in real-time remains a challenge for Intelligent Vehicle Detection Systems (IVDS). Traditional video analysis often struggles with the high computational cost of "dense" optical flow—the pattern of apparent motion calculated for every pixel in a scene. Calculating motion for every pixel in a high-resolution video feed is too slow for real-time enforcement. Furthermore, environmental noise like moving shadows or swaying trees can trigger false positives.
This research addresses these challenges by proposing a sparse optical flow pipeline that focuses on specific regions of interest (ROI) and utilizes reliability verification to distinguish vehicles from background noise.
2. Proposed Methodology
The authors developed a multi-stage pipeline designed to minimize unnecessary calculations and maximize processing speed.
2.1 Lane Detection and ROI Setting
Instead of analyzing the whole frame, the system first identifies the yellow center line. It converts the image from RGB to the HSV color space, which is more robust for isolating specific colors under varying light. It then uses the Progressive Probabilistic Hough Transform (PPHT) to find the straight line of the median. A narrow Region of Interest (ROI) is then drawn around this line to focus the motion analysis.
2.2 Feature Point Selection and Sparse Optical Flow
To further save time, the system uses Harris Corner Detection. Rather than tracking every pixel, it only tracks "corners"—points with high intensity changes in multiple directions—which are more stable and easier to follow across frames. The system applies the Pyramid Lucas-Kanade (PLK) algorithm to these corner points. This "sparse" approach tracks only the selected features within the ROI.
2.3 U-turn Logic and Reliability Verification
The system looks for motion vectors that are perpendicular to the center line (horizontal movement). To filter out noise (like a shadow crossing the line), it uses a reliability algorithm that clusters these vectors. If a group of vectors moves cohesively and exceeds a reliability threshold, it is flagged as an illegal U-turn.
3. Experimental Results
The algorithm was evaluated using real-world footage (640 x 480 resolution) recorded near the Korea National University of Transportation. The proposed method showed a dramatic improvement in speed compared to standard dense optical flow methods:
- Dense Optical Flow (Full Frame): 116.43 ms
- Proposed Sparse PLK (Full Frame): 24.12 ms
- Proposed Sparse PLK (Within ROI): 6.43 ms
The ROI-based sparse method is approximately 18 times faster than traditional dense full-frame analysis. In testing, the reliability verification successfully filtered out non-vehicle motion, resulting in a 0% false detection rate for the provided test samples.
4. Conclusion
By combining color-based lane detection with sparse optical flow focused on a specific ROI, the proposed system achieves high accuracy with a processing time of just 6.43ms per frame. This efficiency makes it a viable solution for real-time traffic enforcement. Future work will explore improving robustness under adverse weather conditions and handling complex occlusion scenarios in dense traffic.
Summary updated