DeepAccident

Deepconcident数据集是第一个用于自动驾驶的大规模事故数据集，该数据集包含全面的传感器集，并支持各种自动驾驶任务。值得注意的是，对于我们设计的每个场景，我们都有四个数据收集工具，其中两个设计用于相互碰撞，另外两个分别跟随。因此，deep事故还可以支持多车辆合作自动驾驶。• 各种事故场景，每种情况下的四种数据收集车辆提供不同的视角，并实现多车辆协作自动驾驶。• 131k带注释的激光雷达样本 (3倍nuScenes) 和791k带注释的带有细粒度对象注释的摄像机图像 (总共六个类别: 汽车，货车，卡车，行人，骑自行车的人，摩托车)。• 支持许多任务: 3D物体检测和跟踪 (基于激光雷达、基于多视图图像和基于单目图像) 、BEV语义和实例分割 (基于多视图图像) 、运动预测。• 各种场景发生的地方、天气、一天中的时间。

The Deepconcident dataset is the first large-scale accident dataset for autonomous driving, which features a comprehensive sensor suite and supports a wide range of autonomous driving tasks. Notably, for each scenario we designed, four data collection vehicles are deployed, with two engineered to collide with each other and the other two tasked with following separately. As such, the Deepconcident dataset also supports multi-vehicle cooperative autonomous driving.
• Diverse accident scenarios: Four data collection vehicles per scenario offer distinct viewpoints, enabling multi-vehicle cooperative autonomous driving.
• 131k annotated LiDAR samples (3x the size of the nuScenes dataset) and 791k annotated camera images with fine-grained object annotations, encompassing six total categories: car, van, truck, pedestrian, cyclist, and motorcycle.
• Supports a comprehensive set of tasks: 3D object detection and tracking (LiDAR-based, multi-view image-based, and monocular image-based), BEV semantic and instance segmentation (multi-view image-based), and motion prediction.
• Diverse scenario settings across varying locations, weather conditions, and times of day.

搜集汇总

背景与挑战

背景概述

DeepAccident是首个用于自动驾驶的大规模事故数据集，提供全面的传感器数据，包括131k激光雷达样本和791k摄像机图像，支持多车辆协作视角和多种任务如3D检测与跟踪。它覆盖多样化的场景、天气和时间条件，专注于事故场景分析，适用于自动驾驶安全研究。