LISA交通灯数据集，超过44分钟的注释交通灯数据

在评估计算机视觉项目时，训练和测试数据是必不可少的。所使用的数据代表了一个被提议的系统将要解决的挑战。我们希望有一个大型的数据库，其中包含大量的变化，例如在城市环境中检测和识别交通信号灯(TL)。从调查现有工作可以清楚地看出，目前的评估主要局限于作者自己收集的小型当地数据集，而不是公开可用的数据集。局部数据集的大小通常很小，并且包含很少的变化。这使得几乎不可能比较不同作者的工作和结果，但也变得很难确定一个领域的当前状态。为了给将来的交通灯识别(TLR)研究提供一个通用的基础，我们收集了一个基于美国道路的视频的广泛的公共数据库。该数据库包括连续测试和训练视频序列，总共43,007帧和113,888个注释交通灯。这些序列是由安装在车顶上的立体摄像机拍摄的，车辆在夜间和白天行驶，光线和天气条件各不相同。该数据库中只使用了左侧摄像机视图，因此立体声特征处于当前状态未使用。数据库收集于美国加利福尼亚州圣地亚哥。该数据库提供了主要用于测试的四个白天和两个夜间序列，在太平洋海滩和拉霍亚提供了23分25秒的驾驶时间。立体图像对是使用点灰色的大黄蜂 XB3(BBX3-13S2C-60)获得的，它包含三个镜头，捕捉图像的分辨率为1280 x 960，每个镜头的视场(FoV)为66 ° 。其中左摄像机视图是用于所有的测试序列和训练剪辑。训练片段包括13个日间片段和5个夜间片段。Zip 包含两种类型的注释

When evaluating computer vision projects, training and test data are indispensable. The data used represents the challenges that a proposed system is intended to solve. We aim to build a large database with extensive variations, such as detecting and recognizing traffic lights (TL) in urban environments. A review of existing work clearly shows that current evaluations are mostly limited to small local datasets collected by the authors themselves, rather than publicly available ones. Local datasets are typically small in scale and contain very limited variations, making it nearly impossible to compare the work and results of different authors, and also difficult to determine the current state of the field. To provide a universal foundation for future traffic light recognition (TLR) research, we collected a comprehensive public database based on videos from U.S. roads. The database includes sequential test and training video sequences, totaling 43,007 frames and 113,888 annotated traffic lights. These sequences were captured by stereo cameras mounted on the vehicle roof, with the vehicle driving during both daytime and nighttime under varying lighting and weather conditions. Only the left camera view was used in this database, so the stereo features remain unused as of the current state. The database was collected in San Diego, California, USA. It provides four daytime and two nighttime sequences primarily for testing, with 23 minutes and 25 seconds of driving footage captured in Pacific Beach and La Jolla. The stereo image pairs were acquired using Point Grey’s Bumblebee XB3 (BBX3-13S2C-60), which is equipped with three lenses, each capturing images at a resolution of 1280 × 960 and a field of view (FoV) of 66°. The left camera view was used for all test sequences and training clips. The training segments include 13 daytime segments and 5 nighttime segments. The Zip archive contains two types of annotations.

搜集汇总

背景与挑战

背景概述

LISA交通灯数据集是一个包含超过44分钟注释交通灯数据的公共数据库，专为计算机视觉中的交通灯识别研究设计。数据集包含43,007帧和113,888个注释交通灯，采集于美国加利福尼亚州圣地亚哥，涵盖多种光线和天气条件下的日间和夜间驾驶视频。