# Human Fall Detection- Multimodality Approach

**Authors:** Xi Wang, Ramya Penta, Bhavya Sehgal, Dale Chen-Song  
**arXiv:** [2302.00224](https://arxiv.org/abs/2302.00224)  
**Date:** 2023-02-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.00224v1/

## Key points

- Frequency of falls among senior citizens has increased, posing a significant risk to their health and well-being
- Introduction of various data sets and machine learning models related to fall detection
- Proposal of a method for human fall detection using a multi-modality approach
- Utilization of UP-FALL detection data set collected by volunteers using different sensors and cameras
- Focus on using wrist sensor data with accelerometer readings and binary labels for fall and no fall
- Fusion of camera and sensor data to enhance performance
- Wrist sensor data alone can be sufficient for accurate fall detection
- Evaluation of model performance using accuracy, precision, recall, and F1 score metrics
- Visualization of evaluation metrics in Figure 7
- Examples of accelerometer readings in Figure 3 and preprocessed images from cameras 1 & 2 in Figure 4 with binary labeling
- Use of multilayer perceptron (MLP) feature extraction and one-dimensional convolutional neural networks (CNNs) for effective analysis of sensor data
- CNNs performed better for unimodal prediction based on sensor data
- Two-dimensional CNN used for feature fusion since image data was available
- Fusion of camera and sensor data did not significantly improve the model's performance in binary classification.

## Summary

In recent years, the frequency of falls among senior citizens has increased, posing a significant risk to their health and well-being. As a result, detecting falls has become crucial, leading to the introduction of various data sets and machine learning models related to fall detection. In this project report titled "Human Fall Detection- Multimodality Approach," the authors propose a method for human fall detection using a multi-modality approach. The authors utilized the UP-FALL detection data set which was collected by numerous volunteers using different sensors and two cameras. Specifically, they focused on using wrist sensor data with accelerometer readings and binary labels for fall and no fall from the data set. To enhance performance, they employed a fusion of camera and sensor data. The experimental results revealed that when compared to using multi-sensor data for binary classification utilizing only wrist data did not significantly impact the model's prediction performance for fall detection. This finding suggests that wrist sensor data alone can be sufficient for accurate fall detection. The authors also evaluated the performance of their model using several evaluation metrics such as accuracy, precision, recall, and F1 score. These metrics provide insights into how well the model performs in correctly identifying true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The visualization of these evaluation metrics is presented in Figure 7. Additionally, the authors provided examples of accelerometer readings showing x, y z axes with FitBit in Figure 3 as well as preprocessed grayscaled 32x32 images from cameras 1 & 2 with binary labeling of falling or not falling in Figure 4. To analyze the sensor data effectively, the authors employed both multilayer perceptron (MLP) feature extraction and one-dimensional convolutional neural networks (CNNs). The results indicated that CNNs performed better for unimodal prediction based on sensor data. Furthermore since image data was also available they used a two-dimensional CNN for feature fusion. In conclusion, the proposed multimodality approach for human fall detection demonstrates that utilizing wrist sensor data alone can yield accurate results however fusion of camera and sensor data did not significantly improve the model's performance in binary classification.