ExoNet Database: Wearable Camera Images of Human Locomotion Environments

Advances in artificial intelligence and robotic vision have enabled researchers to develop environment recognition systems for lower-limb exoskeletons and prostheses. However, insufficient and private training datasets have impeded the widespread development and dissemination of image classification algorithms. To address these shortcomings, we have developed the ExoNet Database, the first open-source database of high-resolution wearable RGB camera images of human locomotion environments. Using a lightweight wearable smartphone camera system, over 5.6 million images of outdoor and indoor real-world walking environments were collected throughout the summer, autumn, and winter seasons. Approximately 940,000 images of the ExoNet Database were human-annotated using a 12-classs hierarchical image classification architecture. Images were uploaded to IEEE DataPort and are publicly available for download. The ExoNet Database provides an unprecedented community platform for training, developing, and comparing next-generation image classification algorithms (e.g., convolutional neural networks) for control of lower-limb exoskeletons and prostheses. Reference: Laschowski B, McNally W, Wong A, and McPhee J. (2020). ExoNet Database: Wearable Camera Images of Human Locomotion Environments. Annual International Conference of the IEEE Engineering in Medicine and Biology Society. In Preparation.

人工智能与机器视觉领域的技术进展，助力研究者开发出面向下肢外骨骼（lower-limb exoskeletons）与假肢（prostheses）的环境识别系统。然而，训练数据集的匮乏与私有属性，阻碍了图像分类算法的大规模开发与推广应用。为解决上述痛点，我们构建了ExoNet数据库——这是首个面向人类步行环境的高分辨率可穿戴RGB（红-绿-蓝）相机图像开源数据集。本研究采用轻量化可穿戴智能手机相机系统，在夏季、秋季与冬季三个季节采集了超过560万张真实室内外步行环境图像。ExoNet数据库中约94万张图像采用12类分层图像分类架构完成人工标注。该数据集已上传至IEEE DataPort平台，可供公众免费下载。ExoNet数据库为训练、开发与对比用于下肢外骨骼及假肢控制的下一代图像分类算法（如卷积神经网络（convolutional neural network, CNN））提供了前所未有的社区共享平台。参考文献：Laschowski B、McNally W、Wong A与McPhee J.（2020）. ExoNet数据库：人类运动环境可穿戴相机图像. IEEE工程医学与生物学学会年度国际会议. 待刊.

搜集汇总

背景与挑战

背景概述

ExoNet是目前最大的开源可穿戴相机行走环境数据集，包含560万张RGB图像和92.3万张人工标注图像，采用12类分层标签架构，覆盖多种季节和场景，为开发视觉智能深度学习模型提供了重要平台。