A pre-impact fall detection data segmentation method based on multi-channel convolutional neural network and class activation mapping

OBJECTIVE

A segmentation method for pre-impact fall detection data is investigated. Specifically, it studies how to partition data segments that are important for classification from continuous inertial sensor data for pre-impact fall detection.

APPROACH

In this study, a trigger-based algorithm combining multi-channel convolutional neural network and class activation mapping was proposed to solve the problem of data segmentation. First, a pre-impact fall detection training dataset was established and divided into two parts. For falls, the 1-second data was divided from the peak value of the acceleration signal magnitude vector to the starting direction. For activities of daily living, the cycle segmentation was performed for a 1-second window size. Second, a heat map of the class activation regions of the sensor data was formed using a multi-channel convolutional neural network and a class activation mapping algorithm. Finally, the data segmentation strategy was established based on the heat map, the basic law of falls and the real-time requirements.

MAIN RESULTS

This method was verified by the SisFall dataset. The obtained segmentation strategy (i.e., to start segmenting a small data segment with a window duration of 325 ms when the acceleration signal magnitude vector is less than 9.217 m/s2) met the real-time requirements for pre-impact fall detection. Moreover, it was suitable for various machine learning algorithms, and the accuracy of the machine learning algorithms used exceeded 94.8%, with the machine learning algorithms verifying the data segmentation strategy.

SIGNIFICANCE

The proposed method can automatically identify the class activation area, save the computing resources of wearable devices, shorten the duration of segmentation window, and ensure the real-time performance of pre-impact fall detection.

Pre-Impact Fall Detection with CNN-Based Class Activation Mapping Method

In this paper, we report our improvement on the prediction accuracy of pre-impact fall detection by applying a learning-based method on the real-time data from an IMU (inertial measurement unit)-sensor mounted on the waist, making it possible to achieve a high accuracy on a wearable device with the extracted features. Using the fixed threshold method is difficult for achieving satisfactory detection accuracy, due to various characteristics and behaviors in the movement of different individuals. In contrast, one could realize high-accuracy detection with machine learning-based methods, but it is difficult to apply them in the wearable devices due to the high hardware requirement. Our method merges the two methods above. We build a convolutional neural network (CNN) with a class activation mapping (CAM) method, which could highlight the class-specific region in the data and obtain a hot map of the fall data. After training on the MobiAct dataset, the model could achieve high-accuracy detection (95.55%) and obtain the region with high contributions to the classification. Then, we manually extract effective features and characteristics of this region and form our special threshold method, achieving pre-impact fall detection in real-world data. Consequently, our method achieves accuracy of 95.33% and a detection time of within 400 ms.