GMDCSA-24: A dataset for human fall detection in videos

The population of older adults (elders) is increasing at a breakneck pace worldwide. This surge presents a significant challenge in providing adequate care for elders due to the scarcity of human caregivers. Unintentional falls of humans are critical health issues, especially for elders. Detecting falls and providing assistance as early as possible is of utmost importance. Researchers worldwide have shown interest in designing a system to detect falls promptly especially by remote monitoring, enabling the timely provision of medical help. The dataset ‘GMDCSA-24′ has been created to support the researchers on this topic to develop models to detect falls and other activities. This dataset was generated in three different natural home setups, where Falls and Activities of Daily Living were performed by four subjects (actors). To bring the versatility, the recordings were done at different times and lighting conditions: during the day when there is ample light and at night when there is low light in addition, the subjects wear different sets of clothes in the dataset. The actions were captured using the low-cost 0.92 Megapixel webcam. The low-resolution video clips make it suitable for use in real-time systems with fewer resources without any compression or processing of the clips. Users can also use this dataset to check the robustness and generalizability of a system for false positives since many ADL clips involve complex activities that may be falsely detected as falls. These complex activities include sleeping, picking up an object from the ground, doing push-ups, etc. The dataset contains 81 falls and 79 ADL video clips performed by four subjects.


a b s t r a c t
The population of older adults (elders) is increasing at a breakneck pace worldwide.This surge presents a significant challenge in providing adequate care for elders due to the scarcity of human caregivers.Unintentional falls of humans are critical health issues, especially for elders.Detecting falls and providing assistance as early as possible is of utmost importance.Researchers worldwide have shown interest in designing a system to detect falls promptly especially by remote monitoring, enabling the timely provision of medical help.The dataset 'GMDCSA-24 has been created to support the researchers on this topic to develop models to detect falls and other activities.This dataset was generated in three different natural home setups, where Falls and Activities of Daily Living were performed by four subjects (actors).To bring the versatility, the recordings were done at different times and lighting conditions: during the day when there is ample light and at night when there is low light in addition, the subjects wear different sets of clothes in the dataset.The

Value of the Data
• Human falls, a significant health concern especially for elders, require early detection, emphasizing the importance of an efficient and accurate fall detection system [ 1 , 2 ].So, researchers are increasingly interested in developing and implementing an efficient and accurate human fall detection system.This dataset can be used to train or test a human fall detection system.• The video recording in the proposed dataset was captured in three different homes and multiple environments, including varying lighting conditions, by four subjects wearing different attires.• Besides falls, the dataset contains many activities of daily living (ADL).Therefore, it also can be used for human activity recognition (HAR) [ 3 , 4 ].• The dataset was generated using a low-resolution (0.92 MP) webcam, making it computationally efficient without further compressing or processing.That is, this raw data is suitable for real-time use on low-computing devices [ 1 , 5 , 6 ]. • Different occlusions are incorporated into the dataset to enrich its diversity.
• Some ADL videos in this dataset feature activities similar to falls, such as sleeping, doing push-ups, etc.These activities closely resemble falls.So, one of the goals of this dataset is to assess the robustness of fall detection systems in handling false positives.

Background
According to a report by the United Nations Population Fund (UNFPA), the average lifespan has increased globally from 45.51 years in 1950 to 73.16 years in 2023 [ 7 ].Simultaneously, the average fertility rate has decreased from 5 in 1950 to 2.3 in 2021 on a global scale [ 7 ].This is causing an imbalance where there are more aged persons than younger people, so support for the elders has to be largely increased, especially for their independent living.Falling is a common but potentially devastating experience for elders.If immediate medical care is not provided, it can be fatal or lead to disability [ 8 ].Therefore, there is a need for indoor automated systems that can monitor and detect falls in elders in their homes.In this data article, we present a human fall dataset named 'GMDCSA-24 to assist researchers in developing models for detecting human falls and other non-fall activities (ADL).Some highlights of the released dataset are represented below.
• Though there are many human fall datasets [9][10][11][12][13][14][15][16][17], not all of them are easily accessible [ 11,14,16 ]. • The recording of many fall datasets has been done in a single home environment [ 12 , 13 , 15 , 18 ], unlike the proposed dataset, which has been recorded in three natural home environments.This diversity in environments provides a broader range of scenarios and variations, which can improve the robustness and generalizability of fall detection models trained on this dataset, leading to more accurate and reliable predictions in real-world applications.• Many fall datasets [ 10-12 , 14-18 ] do not include data with occlusions, unlike our proposed dataset, which contains many frames where subjects are occluded.Including occlusions provides a more realistic representation of real-world conditions, improving the model's ability to handle challenging scenarios and increasing the robustness and accuracy of the models.• In one of the released datasets [ 9 ], only one actor is depicted, while in some datasets [ 16 , 17 ], the total number of subjects is not exactly provided by the dataset creators, unlike the proposed dataset, which includes four subjects.Having more subjects increases the diversity of the dataset, which helps fall detection models generalize better across different individuals.This diversity leads to improved model performance and reliability when applied to varied real-world scenarios.• The total number of activities (fall and ADL) performed in many datasets is less than in the proposed dataset, and for some, it's not provided by the dataset creators [ 10 , 12-14 , 18 ].Having a greater number of activities increases the variety and complexity of the dataset, which helps machine learning models learn more comprehensive patterns.This leads to enhanced model accuracy and robustness when applied to diverse and complex real-world scenarios.• The total storage size of the proposed dataset is also less than many available datasets [ 9 , 10 , 12 , 15 , 17 , 18 ].A smaller storage size makes the dataset more manageable and accessible, facilitating faster data processing and reducing computational resource requirements.This can lead to quicker experimentation and model training, making it easier to deploy models in resource-constrained environments.• Many existing datasets are of very high quality, making them unsuitable to use directly (without compressing or further processing) on low-resource computing devices.
Table 1 compares the proposed GMDCSA-24 dataset with the existing human fall dataset.

Data Description
The 'GMDCSA-24 dataset is an extension of the 'GMDCSA' dataset [ 1 , 5 ].The original GMD-CSA human fall dataset was created by performing fall and ADL activities with a single subject consisting of 16 fall and 16 ADL clips.The GMDCSA-24 dataset includes fall and ADL activities video clips using three additional actors in two new home setups.Recordings were made at  different times of the day, with varying levels of lighting creating some challenges for model development in fall detection.As mentioned in the 'Value of the Data' section, the GMDCSA-24 dataset is unique in its use of a low-megapixel (0.92 MP) integrated laptop webcam, making it suitable to use directly (without any compression) on resource-constrained devices [ 19 , 20 ].Another advantage of the GMDCSA-24 dataset is the inclusion of fall-like activities such as sleeping and doing push-ups, etc. in the ADL videos.This similarity between sleeping and falling poses a challenge for fall detection pipelines, as they may incorrectly classify sleep as a fall event.So, this dataset is valuable for testing the robustness of models in reducing false positives.2 .
The CSV files describe the file name, length in seconds, time of recording, attire, description, and classes, as shown in Tables 3 -and 5 .The start and end times for each class are indicated in square brackets in seconds.To separate words within a field, a semicolon (';') is used instead of a comma (','), so it is not treated as a new field.The classes are listed in alphabetical order for ADL, and for falls, the fall classes are listed first in alphabetical order, followed by the ADL classes in alphabetical order.If a class appears multiple times in a clip, its timings are separated by a semicolon (';'), as shown in Table 4 .
The subject and class-wise details of the length and dimensions of the clips of this dataset are shown in Table 6 .There are four subjects from Subject 1 to Subject 4. The length column displays the minimum, maximum, mode, median, and mean value of the clip duration in seconds.The Dimension column displays the two types of dimensions of the video clips of this dataset, along with the number of clips for each dimension.
Tables 7 and 8 provide a brief description of the ADL and fall activities performed by Subject 2. In the same manner, Tables 9 and 10 describe the ADL and fall activities of Subject 3. Similarly, Tables 11 and 12 outline the ADL and fall activities of Subject 4.  Tables 13-20 list the individual activities that appear in each file in the Subject 1, Subject 2, Subject 3, and Subject 4 directories.This information can be useful for the task of HAR.We have only mentioned common activities like drinking, eating, exercising, reading, sitting, sleeping, standing, walking, and writing for the ADL class.Fall backward (BW), fall forward (FW), and fall sideways (SW) are also mentioned for the Fall class.
Table 21 summarizes all the activities and their frequencies by the four subjects of the GMDCSA-24 dataset.The rows of Table 21 are ordered alphabetically, first by ADL and then by fall activities.
Figs. 2 and 3 show some sample frames from the ADL and fall video clips, respectively, from Subject 1, Subject 2, Subject 3, and Subject 4. The file names of each frame are shown below each image.

Experimental Design, Materials and Methods
This dataset was created by capturing the fall and ADL activities performed by four different subjects in three different home setups.The subjects were asked to perform random, natural, and common ADL and fall activities in indoor setups, wearing different sets of clothes and recording at different times of the day.This makes this dataset very suitable and versatile for any fall detection models.All subjects were informed about the use of this dataset, and consent was obtained from them.This dataset incorporates numerous ADL video sequences that closely resemble falls, featuring actions such as i) sleeping, ii) picking something up from the ground, iii) exercises similar to falls, like push-ups, etc.One of the primary uses of this dataset is to assess the system's robustness in detecting false positives.Additionally, the dataset boasts a lower resolution than other datasets, facilitating swift training and testing times without any further compression.
The camera (laptop) was kept fixed (static) while capturing the activities performed by all four subjects.As mentioned in the Data Description section, the video clips were captured using the 0.92 MP (720p, 30 FPS) webcam of the HP G5 348 laptop (Intel Core i5 8th Generation).The LosslessCut1 software was used to trim some lengthy video clips.The VLC media player2 was used to play back and check the videos.To record the videos, the Microsoft Camera (version 2024.2405.19.0) app3 was used.The classification of different ADL and fall activities was done manually after playing the video clips.The CSV files were prepared manually after playing the clips.The VLC extension, Time v3.2, 4 was used to see the precise playback time.
Fall and ADL video clips were recorded in three natural room setups.Subject 1 performed in Room 1, Subject 2 and Subject 3 in Room 2, and Subject 4 in Room 3. The details of the rooms are provided in Table 22 .Two camera positions were used in Room 1: one from the gate side towards the bed at a height of 70 cm and another from the bed towards the gate at a height of 90 cm.For Room 2 and Room 3, a fixed camera position (towards the bed) was used.

Fig. 1
illustrates the organization of this dataset.The GMDCSA-24 dataset comprises four subdirectories: Subject 1, Subject 2, Subject 3, and Subject 4. Each directory, Subject 1, Subject 2, Subject 3, and Subject 4, contains two subdirectories: ADL and Fall, along with two CSV files, ADL.csv and Fall.csv.Each ADL and Fall directory contains video clips in MP4 format.Both the ADL and Fall directories under Subject 1 contain 16 video clips, resulting in a total of 32 video clips.Similarly, Subject 2 has 23 clips in the ADL directory and 25 clips in the Fall directory, while Subject 3 contains 22 ADL clips and 21 fall clips.Subject 4 contains 20 ADL clips and 17 fall clips.The basic details of the video clips of this dataset are shown in Table

Falling backward on theFig. 2 .
Fig. 2. Some sample frames from the ADL class of the GMDCSA-24 dataset from all four subjects.

Fig. 3 .
Fig. 3. Some sample frames from the Fall class of the GMDCSA dataset from all four subjects.

Table 1
Comparison of the existing vision-based human fall datasets based on size, accessibility, home environment, occlusion, No. of subjects, No. of videos, year, etc., with the proposed dataset.
a As of 21th July 2024.b Including other supporting fillies.

Table 3
Sample value of the ADL.csv file of the Subject 1 directory.

Table 4
Sample value of the ADL.csv file of the Subject 3 directory where there are multiple timings for a class.

Table 5
Sample value of the Fall.csvfile of the Subject 2 directory.

Table 6
Attributes of the ADL and Fall videos.

Table 7
ADL class files descriptions of Subject 2.

Table 8
Fall class files descriptions of Subject 2.

Table 9
ADL class files descriptions of Subject 3.

Table 10
Fall class files descriptions of Subject 3.