Multi-modal Fall Detection via Sensor-Video Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fall detection technologies, both wearable sensor-based and video data-based, face challenges with insufficient detection accuracy, especially in demanding situations, and fail to provide accurate location information, leading to delayed assistance.
Innovation Solution
A system that combines sensor data and video data to analyze signal features indicative of a potential fall, correlating changes over time between sensor and video data streams to enhance detection accuracy and provide precise location information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fall detection uses only wearable sensors, then the device can provide location information, but detection accuracy is insufficient in demanding situations
Solution Approach 1:
The patent combines wearable sensor data (accelerometers, gyroscopes) with video camera data to create a multi-modal fall detection system. The analysis unit processes both sensor signals and video frames simultaneously, comparing results from both sources to confirm fall events. This merging of different sensing modalities resolves the contradiction by maintaining detection accuracy while improving reliability in demanding situations through cross-validation of fall detection from multiple independent sources.
2Measurement precision
If fall detection uses only video data, then location information is provided, but detection accuracy is poor in realistic usage conditions
Solution Approach 1:
The system merges video-based fall detection with wearable sensor-based detection. The analysis unit correlates video frame analysis results with sensor data patterns, using the sensor data as a trigger to initiate more intensive video analysis. This combination maintains the location-providing capability of video while achieving high detection accuracy through the complementary strengths of both modalities, resolving the contradiction between accuracy and reliability in realistic conditions.
3Measurement precision
If the system combines sensor data and video data for analysis, then detection accuracy is improved, but device complexity increases
Solution Approach 1:
The system uses wearable sensors to perform preliminary fall detection and triggers video analysis only when sensor data indicates a potential fall event. This preliminary action by the sensors filters out normal activities, allowing the more complex video analysis to focus only on suspicious cases. This resolves the contradiction by maintaining high detection accuracy through combined analysis while reducing overall system complexity and computational burden through selective activation of video processing.
Solution Approach 2:
The system dynamically adjusts its analysis mode based on sensor input. During normal conditions, it operates in low-power sensor monitoring mode. When sensor data indicates a potential fall, it dynamically switches to intensive video analysis mode. This dynamic operation resolves the contradiction by providing high accuracy when needed while minimizing complexity and power consumption during normal operation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a device, system and method for reliable and fast fall detection. The device comprises a sensor input (11) for obtaining sensor data related to movement of a subject acquired by a body-worn sensor (20, 22, 61) worn by the subject, and a video input (12) for obtaining video data of the subject and/or the subject's environment. An analysis unit (13) analyzes the obtained sensor data to detect one or more signal features indicative of a potential fall of the subject and for analyzing, if one or more such signal features have been detected, the detected one or more signal features and/or related sensor data in combination with related video data to identify similarities between: changes of the related sensor data and/or the signal features over time; and changes of the video data over time. An output (14, 64, 65, 66) issues a fall detection indication if the level and/or amount of detected similarities exceeds a corresponding threshold.