2025年动态环境多视角高帧率视觉感知数据集

2025年动态环境多视角高帧率视觉感知数据集（Spharx 7H45 AI Visual Perception Dataset）简称7H45数据集，是一个应用于人工智能通用感知训练的多模态数据集。该数据集由杭州极光感知科技有限公司基于自研设备、自研算法的自采数据构建，以RAW视频流取帧格式为主，传感器全画幅图像帧为辅。该数据集应用聚焦于两大方向：1.纯视觉方案完全自动驾驶的感知算法开发与验证，为高速行驶实现精准的目标检测、密集场景语义分割、距离计算及动态目标行为预测等任务提供关键数据；2.作为具身智能机器人的视觉能力“预训练基座”，所蕴含的丰富动态信息、复杂光照变化及强运动模糊特性，能有效训练出对现实世界具有高鲁棒性的通用视觉表征，通过迁移学习可适配至各类机器人平台，极大提升在真实非结构化环境中进行导航、避障与人机交互的底层感知能力。The 2025 Dynamic Environment Multi-View High-Frame-Rate Visual Perception Dataset (Spharx 7H45 AI Visual Perception Dataset), abbreviated as the 7H45 Dataset, is a multimodal dataset intended for general artificial intelligence perception training. Constructed by Hangzhou Aurora Perception Technology Co., Ltd. from self-collected data using self-developed equipment and algorithms, this dataset primarily adopts RAW video stream frame extraction format, supplemented by full-frame sensor image frames. Its applications focus on two core directions: 1. Development and validation of perception algorithms for fully autonomous driving with pure vision-based solutions, providing critical data for tasks including accurate object detection, dense scene semantic segmentation, distance estimation, and dynamic target behavior prediction to support high-speed driving scenarios; 2. Serving as a pretraining backbone for the visual capabilities of embodied intelligent robots. The dataset contains abundant dynamic information, complex lighting variations, and severe motion blur characteristics, which can effectively train general visual representations with high robustness to real-world environments. Through transfer learning, these representations can be adapted to various robotic platforms, greatly enhancing the underlying perception capabilities for navigation, obstacle avoidance, and human-robot interaction in real unstructured environments.

搜集汇总

背景与挑战

背景概述

该数据集是由极光感知科技自主研发的多模态视觉数据集，主要用于自动驾驶感知算法开发和机器人视觉预训练，其高帧率RAW视频流和复杂动态场景数据可提升AI系统在真实环境中的鲁棒性表现。