跌落检测数据集，用于跌倒检测系统

跌倒检测系统的目标是自动检测人类跌倒并可能已经受伤的情况。这种系统的一个自然应用是对病人和老人的家庭监控，以便在跌倒造成伤害的情况下自动提醒亲属和/或当局。我们提出了一种基于Kinect的统计方法，通过考虑人跌倒的帧数、跌倒的幅度、跌倒的最大速度和跌倒过程中逐帧下降的速度，根据最后几帧做出判断。由于深度传感器的范围是0:5米到4米，一个Kinect不足以覆盖整个空间。我们在家庭环境中设置了两个Kinects。我们独立于用户和相机的测试表明，我们的方法在现实生活中是适用的。 实验场景 本文的实验数据来自于德克萨斯大学阿灵顿分校Heracleia以人为本计算实验室的实验。在这个实验室里，设置了一个模拟的公寓。两个Kinects被设置在公寓的两个角落，并被设置为监视公寓。设置两个Kinects的原因是深度传感器的范围是从0:5米到4米，这意味着一个Kinect不足以覆盖整个公寓。下图中第一行是视图1，第二行是视图2，左边是深度图，右边是颜色图。 实验场景 本文的实验数据来自德克萨斯大学阿灵顿分校的 Heracleia人类中心计算实验室的实验。 在这个实验室中，已经设置了一个模拟公寓。在公寓的两个角落设置了两个 Kinect，并设置了监控公寓。设置两个 Kinect 的原因是深度传感器的范围为 0：5m 到 4m，这意味着一个 Kinect不足以覆盖整个公寓。 左侧为深度图，右侧为彩色地图。 实验数据 六个被试在两个场景中分别做了几个动作。 这些动作包括真正的跌倒和其他类似跌倒的动作，如从地上捡起硬币，在地上坐下，系鞋带等。场景1中有10400帧和12个真实摔倒，而场景2中有21214帧和14个真实摔倒。下表显示了我们实验中的类似摔倒的动作。 在上表中，pf表示从地板上捡东西，ts表示系鞋带，sb表示在床上睡觉，sif表示坐在地板上，pd表示打开下面的抽屉，这个抽屉离地板很近，jb表示跳到地上，sf表示在地板上睡觉。下面的数字显示了一个跌倒的过程。 A fall process 在注释文件中，格式是这样的。Alexis view1 202 215。Alexis是用户名。view1表示场景1。202是开始帧，215是结束帧。下面的数字显示了一个典型的类似摔倒的动作，就是坐在地上。 Sitting on the floor

The goal of a fall detection system is to automatically identify instances where a human has fallen and may have suffered injuries. A typical application of such a system is home monitoring for patients and the elderly, to automatically notify relatives and/or relevant authorities when a fall results in injury. We propose a Kinect-based statistical approach that makes judgments based on the most recent several frames, by taking into account the number of frames during the fall, the fall amplitude, the maximum fall speed, and the per-frame descent speed throughout the fall sequence. Given that the operating range of a depth sensor is 0.5 meters to 4 meters, a single Kinect cannot cover an entire space. We deployed two Kinect sensors in a home environment. Evaluations independent of specific users and camera setups demonstrate that our method is applicable in real-world scenarios.

Experimental Scenarios
The experimental data in this paper are collected from experiments conducted at the Heracleia Human-Centered Computing Laboratory, The University of Texas at Arlington. A simulated apartment was established in this laboratory. Two Kinect sensors were installed at two corners of the apartment to monitor the entire space. The rationale for using two Kinects is that the depth sensor has a range of 0.5m to 4m, meaning a single unit cannot cover the entire apartment. As shown in the figure below, the first row corresponds to View 1, the second row corresponds to View 2; the left side displays the depth map, and the right side displays the color map.

Experimental Data
Six subjects performed multiple actions across two scenarios. These actions include real falls and fall-like motions, such as picking up coins from the floor, sitting on the floor, tying shoelaces, and so on. Scenario 1 contains 10,400 frames and 12 real fall incidents, while Scenario 2 contains 21,214 frames and 14 real fall incidents. The table below lists the fall-like actions included in our experiment.
In the aforementioned table, pf denotes picking up objects from the floor, ts stands for tying shoelaces, sb represents sleeping on the bed, sif refers to sitting on the floor, pd indicates opening a lower drawer that is close to the ground, jb means jumping onto the floor, and sf stands for sleeping on the floor. The figures below demonstrate a fall process.

A Fall Process
The format in the annotation file is as follows: "Alexis view1 202 215". Here, Alexis is the username, view1 indicates Scenario 1, 202 is the start frame, and 215 is the end frame. The figures below show a typical fall-like action: sitting on the floor.

Sitting on the Floor

搜集汇总

背景与挑战

背景概述

该数据集是一个用于跌倒检测系统的数据集，特别设计用于病人和老人的家庭监控。数据集包含来自两个Kinect传感器的数据，覆盖了真实的跌倒动作和类似跌倒的动作，旨在提高跌倒检测的准确性。