PVDN (Provident Vehicle Detection at Night)

对于高级驾驶辅助系统，尽早获得有关迎面而来的车辆的信息至关重要。在夜间，由于照明条件差，这项任务尤其困难。为此，在夜间，每辆车都使用前照灯来改善视线，从而确保安全驾驶。作为人类，我们通过检测由前照灯引起的光反射，在车辆实际可见之前直觉地假设迎面而来的车辆。有了这个数据集，我们提供了一个数据集，其中包含 59746 个注释灰度图像，这些图像来自夜间农村环境中的 346 个不同场景。在这些图像中，所有迎面而来的车辆、它们对应的灯光对象（例如，前照灯）以及它们各自的光反射（例如，护栏上的光反射）都被标记了。有了这个，我们将提供第一个包含综合地面实况数据的开源数据集，以便研究基于它们引起的光反射检测迎面而来的车辆的新方法，远在它们直接可见之前。我们认为这是进一步缩小当前高级驾驶员辅助系统与人类行为之间的性能差距的重要一步。

For Advanced Driver Assistance Systems (ADAS), obtaining information about oncoming vehicles as early as possible is critically important. This task is particularly challenging at night due to poor lighting conditions. To address this issue, all vehicles use headlights at night to improve visibility and ensure safe driving. As humans, we intuitively infer the presence of an oncoming vehicle before it is actually visible by detecting light reflections caused by its headlights.

With this dataset, we present an open-source resource containing 59,746 annotated grayscale images captured from 346 distinct scenes in nighttime rural environments. In these images, all oncoming vehicles, their corresponding light-emitting objects (e.g., headlights), and their respective light reflections (e.g., light reflections on guardrails) have been fully labeled.

This work provides the first open-source dataset with comprehensive ground truth data to support research on novel methods for detecting oncoming vehicles via the light reflections they induce, long before the vehicles themselves become directly visible. We believe this represents a critical step toward further narrowing the performance gap between current Advanced Driver Assistance Systems and human driving behaviors.

搜集汇总

背景与挑战

背景概述

PVDN数据集是一个专注于夜间预知车辆检测的开源数据集，包含59746个灰度图像，来自346个夜间农村场景，标注了车辆、前照灯及光反射对象。该数据集旨在通过光反射检测车辆，以在车辆直接可见之前进行预测，从而提升高级驾驶辅助系统在夜间的性能，适用于计算机视觉和自动驾驶研究。