Abstract
This paper presents a pragmatic approach to real-time vehicle detection for autonomous driving, balancing environmental robustness with computational efficiency. By fusing knowledge-based features—specifically shadows and horizontal edge "waves"—with a temporal motion trajectory verification step, the system achieves high performance without the need for heavy offline-trained classifiers. Quantitative analysis on public datasets and real-time video demonstrates a detection rate of 85.58% with a 4.13% false positive rate, operating at 40ms per frame (25 FPS) on a standard PC platform.
Introduction
Vehicle detection is a fundamental component of Forward Collision Warning (FCW) and Blind Spot Detection (BSD) systems. Existing methods typically fall into two categories: knowledge-based methods, which are fast but prone to false positives from environmental noise, and appearance-based methods, which are accurate but computationally intensive and limited by training data diversity. This research bridges the gap by combining low-cost physical feature extraction with online motion tracking to achieve reliable, real-time on-board detection.
Methodology
The proposed system operates in two distinct phases: Hypothesis Generation (HG) and Hypothesis Verification (HV).
Hypothesis Generation (HG)
The HG phase identifies potential vehicle locations (Regions of Interest, ROIs) using two primary features:
- Shadow Detection: Utilizing adaptive grayscale histograms, the system identifies the dark regions typically found under vehicles on asphalt roads.
- Vehicle Wave (HEPH): The "Horizontal Edge Pixels Histogram" (HEPH) captures the dense horizontal line structures characteristic of vehicle fronts and rears. A horizontal Sobel filter is applied to generate a histogram of these edges, creating a "wave" signature.
The fusion of these features allows the system to quickly narrow down candidate ROIs without scanning the entire image frame.
Hypothesis Verification (HV)
The core innovation lies in the verification step. Instead of a complex neural network, the system uses Motion Trajectory. By tracking ROIs across continuous frames, the algorithm filters out "discrete" noise—objects that appear momentarily or move inconsistently. Real vehicles maintain a stable relative velocity and trajectory, allowing the system to discard transient false positives like road markings or lighting flickers.
Experimental Results
The system was evaluated using the Caltech 1999 dataset and a custom video dataset covering various weather conditions (sunny, rainy, snowy).
- Efficiency: The algorithm processes frames in approximately 40ms (25 FPS), making it suitable for real-time on-board applications.
- Accuracy: The average detection rate (DR) across video tests was 85.58%.
- Robustness: The motion trajectory filter significantly reduced false alarms. In one test scenario, the false positive rate dropped from 51% to just 5% after applying temporal verification.
While highly effective, the system faces challenges in high-density traffic and during heavy rain, where occlusions can disrupt the shadow and edge features.
Conclusion
By treating time as a dimension for verification, this approach allows for a "lightweight" initial detection phase that remains robust against environmental noise. Future work will focus on improving multi-vehicle tracking and handling occlusions to further enhance autonomous driving capabilities.
Summary updated