This paper explores the personalization of smartwatch-based fall detection models trained using a combination of deep neural networks with ensemble techniques. Deep neural networks face practical challenges when used for fall detection, which in general tend to have limited training samples and imbalanced datasets. Moreover, many motions generated by a wrist-worn watch can be mistaken for a fall. Obtaining a large amount of real-world labeled fall data is impossible as fall is a rare event. However, it is easy to collect a large number of non-fall data samples from users. In this paper, we aim to mitigate the scarcity of training data in fall detection by first training a generic deep learning ensemble model, optimized for high recall, and then enhancing the precision of the model, by collecting personalized false positive samples from individual users, via feedback from the SmartFall App. We performed real-world experiments with five volunteers and concluded that a personalized fall detection model significantly outperforms generic fall detection models, especially in terms of precision. We further validated the performance of personalization by using a new metric for evaluating the accuracy of the model via normalizing false positive rates with regard to the number of spikes of acceleration over time.
more »
« less
Transfer Learning on Small Datasets for Improved Fall Detection
Falls in the elderly are associated with significant morbidity and mortality. While numerous fall detection devices incorporating AI and machine learning algorithms have been developed, no known smartwatch-based system has been used successfully in real-time to detect falls for elderly persons. We have developed and deployed a SmartFall system on a commodity-based smartwatch which has been trialled by nine elderly participants. The system, while being usable and welcomed by the participants in our trials, has two serious limitations. The first limitation is the inability to collect a large amount of personalized data for training. When the fall detection model, which is trained with insufficient data, is used in the real world, it generates a large amount of false positives. The second limitation is the model drift problem. This means an accurate model trained using data collected with a specific device performs sub-par when used in another device. Therefore, building one model for each type of device/watch is not a scalable approach for developing smartwatch-based fall detection system. To tackle those issues, we first collected three datasets including accelerometer data for fall detection problem from different devices: the Microsoft watch (MSBAND), the Huawei watch, and the meta-sensor device. After that, a transfer learning strategy was applied to first explore the use of transfer learning to overcome the small dataset training problem for fall detection. We also demonstrated the use of transfer learning to generalize the model across the heterogeneous devices. Our preliminary experiments demonstrate the effectiveness of transfer learning for improving fall detection, achieving an F1 score higher by over 10% on average, an AUC higher by over 0.15 on average, and a smaller false positive prediction rate than the non-transfer learning approach across various datasets collected using different devices with different hardware specifications.
more »
« less
- NSF-PAR ID:
- 10434787
- Date Published:
- Journal Name:
- Sensors
- Volume:
- 23
- Issue:
- 3
- ISSN:
- 1424-8220
- Page Range / eLocation ID:
- 1105
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The majority of current smart health applications are deployed on a smartphone paired with a smartwatch. The phone is used as the computation platform or the gateway for connecting to the cloud while the watch is used mainly as the data sensing device. In the case of fall detection applications for older adults, this kind of setup is not very practical since it requires users to always keep their phones in proximity while doing the daily chores. When a person falls, in a moment of panic, it might be difficult to locate the phone in order to interact with the Fall Detection App for the purpose of indicating whether they are fine or need help. This paper demonstrates the feasibility of running a real-time personalized deep-learning-based fall detection system on a smartwatch device using a collaborative edge-cloud framework. In particular, we present the software architecture we used for the collaborative framework, demonstrate how we automate the fall detection pipeline, design an appropriate UI on the small screen of the watch, and implement strategies for the continuous data collection and automation of the personalization process with the limited computational and storage resources of a smartwatch. We also present the usability of such a system with nine real-world older adult participants.more » « less
-
more » « less
Real-time fall detection using a wearable sensor remains a challenging problem due to high gait variability. Furthermore, finding the type of sensor to use and the optimal location of the sensors are also essential factors for real-time fall-detection systems. This work presents real-time fall-detection methods using deep learning models. Early detection of falls, followed by pneumatic protection, is one of the most effective means of ensuring the safety of the elderly. First, we developed and compared different data-segmentation techniques for sliding windows. Next, we implemented various techniques to balance the datasets because collecting fall datasets in the real-time setting has an imbalanced nature. Moreover, we designed a deep learning model that combines a convolution-based feature extractor and deep neural network blocks, the LSTM block, and the transformer encoder block, followed by a position-wise feedforward layer. We found that combining the input sequence with the convolution-learned features of different kernels tends to increase the performance of the fall-detection model. Last, we analyzed that the sensor signals collected by both accelerometer and gyroscope sensors can be leveraged to develop an effective classifier that can accurately detect falls, especially differentiating falls from near-falls. Furthermore, we also used data from sixteen different body parts and compared them to determine the better sensor position for fall-detection methods. We found that the shank is the optimal position for placing our sensors, with an F1 score of 0.97, and this could help other researchers collect high-quality fall datasets.
-
more » « less
Abstract Advances in visual perceptual tasks have been mainly driven by the amount, and types, of annotations of large-scale datasets. Researchers have focused on fully-supervised settings to train models using offline epoch-based schemes. Despite the evident advancements, limitations and cost of manually annotated datasets have hindered further development for event perceptual tasks, such as detection and localization of objects and events in videos. The problem is more apparent in zoological applications due to the scarcity of annotations and length of videos-most videos are at most ten minutes long. Inspired by cognitive theories, we present a self-supervised perceptual prediction framework to tackle the problem of temporal event segmentation by building a stable representation of event-related objects. The approach is simple but effective. We rely on LSTM predictions of high-level features computed by a standard deep learning backbone. For spatial segmentation, the stable representation of the object is used by an attention mechanism to filter the input features before the prediction step. The self-learned attention maps effectively localize the object as a side effect of perceptual prediction. We demonstrate our approach on long videos from continuous wildlife video monitoring, spanning multiple days at 25 FPS. We aim to facilitate automated ethogramming by detecting and localizing events without the need for labels. Our approach is trained in an online manner on streaming input and requires only a single pass through the video, with no separate training set. Given the lack of long and realistic (includes real-world challenges) datasets, we introduce a new wildlife video dataset–nest monitoring of the Kagu (a flightless bird from New Caledonia)–to benchmark our approach. Our dataset features a video from 10 days (over 23 million frames) of continuous monitoring of the Kagu in its natural habitat. We annotate every frame with bounding boxes and event labels. Additionally, each frame is annotated with time-of-day and illumination conditions. We will make the dataset, which is the first of its kind, and the code available to the research community. We find that the approach significantly outperforms other self-supervised, traditional (e.g., Optical Flow, Background Subtraction) and NN-based (e.g., PA-DPC, DINO, iBOT), baselines and performs on par with supervised boundary detection approaches (i.e., PC). At a recall rate of 80%, our best performing model detects one false positive activity every 50 min of training. On average, we at least double the performance of self-supervised approaches for spatial segmentation. Additionally, we show that our approach is robust to various environmental conditions (e.g., moving shadows). We also benchmark the framework on other datasets (i.e., Kinetics-GEBD, TAPOS) from different domains to demonstrate its generalizability. The data and code are available on our project page:
https://aix.eng.usf.edu/research_automated_ethogramming.html -
Falls are the second leading cause of accidental or unintentional injuries/deaths worldwide. Accurate pose estimation using commodity mobile devices will help early detection and injury assessment of falls, which are essential for the first aid of elderly falls. By following the definition of fall, we propose a P ervasive P ose Est imation scheme for fall detection ( P \( ^2 \) Est ), which measures changes in tilt angle and height of the human body. For the tilt measurement, P \( ^2 \) Est leverages the pointing of the mobile device, e.g., the smartphone, when unlocking to associate the Device coordinate system with the World coordinate system. For the height measurement, P \( ^2 \) Est exploits the fact that the person’s height remains unchanged while walking to calibrate the pressure difference between the device and the floor. We have prototyped and tested P \( ^2 \) Est in various situations and environments. Our extensive experimental results have demonstrated that P \( ^2 \) Est can track the body orientation irrespective of which pocket the phone is placed in. More importantly, it enables the phone’s barometer to detect falls in various environments with decimeter-level accuracy.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.3390/s23031105
