Enhanced nighttime vehicle detection for on-board processing

Resumen

Nighttime vehicle detection poses significant challenges, particularly in scenarios with limited lighting, where visibility is often compromised. To address this problem, this paper proposes a novel nighttime vehicle detection system that dynamically adapts to extreme lighting conditions, ranging from bright daytime scenarios to challenging nighttime conditions where the vehicle's appearance may be entirely lost. For this purpose, a multi-granularity detection approach is adopted, automatically combining bounding-box and point-based representations depending on the vehicle's visibility. Bounding-box detections, reporting location and size information, are selected when the vehicle appearance is mostly visible, such as in daytime or urban nighttime scenarios with sufficient artificial street illumination. Point-based detections, indicating only location information, are used when the vehicle's appearance is not discernible, such as in rural nighttime scenarios with little or no street illumination. The system is designed as a multi-head neural network built on a shared Hourglass backbone that accepts bounding-box and point-based annotations for training and can automatically predict, depending on the scenario, vehicle bounding boxes or point-based predictions. Extensive evaluations on a combined dataset of BDD100K and PVDN demonstrate that the proposed system achieves higher detection accuracy and robustness compared to existing methods, with mean Average Precision (mAP) scores of 0.7134 on BDD100K, 0.6621 on PVDN, and 0.6814 on the combined dataset. Additionally, a self-acquired dataset, FNTVD, further enhances the evaluation by providing real-world driving conditions. The system also achieves real-time performance at 45.45 FPS, making it suitable for practical applications.