A Framework for Anomaly Identification Applied on Fall Detection

Abstract

Automatic systems to monitor people and subsequently improve people's lives have been emerging in the last few years, and currently, they are capable of identifying many activities of daily living (ADLs). An important field of research in this context is the monitoring of health risks and the identification of falls. It is estimated that every year, one in three persons older than 65 years will fall, and fall events are associated with high mortality rates among the elderly. We propose an anomaly identification framework to detect falls, which incorporates a spatial-temporal convolutional graph network (ST-GCN) as a feature extractor and uses an encoder process to reconstruct ADLs and identify falls as anomalies. As the publicly available fall datasets are few and generally unbalanced, training a reliable model using approaches that need explicit labeling is challenging. Thus, a focus on learning without external supervision is desirable. Treating a fall as an exception of ADLs allows us to recognize falls as anomalies without explicit labels. Given its modular architecture, our framework can robustly represent visual information and use the encoder's reconstruction error to identify falls as anomalies. We assess our framework's ability to recognize falls by training it with only ADLs. We perform three types of experiments: single dataset training and evaluation that consists of separate 90% of the data to train the model 5% to adjust the model, and the rest to the test. A joint dataset experiment, where we combine two datasets to increase the number of samples our model is trained on, and a cross-dataset evaluation, where we train on one dataset and evaluate using another one. Besides presenting state-of-the-art results on our experiments, particularly on the cross-dataset one, the model also presents a low number of false events, which makes it an ideal candidate for real-world application.

Human Activity Recognition by Sequences of Skeleton Features

“…Table 4 compares our performance to other studies performing fall detection with camera view skeleton sequences for fall detection and using the same database: UP-Fall. It is possible to observe that the performance of our proposal exceeds the performance delivered by the work carried out in [ 12 , 17 , 32 ]. Moreover, unfortunately, the works in [ 17 , 32 ] do not implement activity recognition.…”

“…It is possible to observe that the performance of our proposal exceeds the performance delivered by the work carried out in [ 12 , 17 , 32 ]. Moreover, unfortunately, the works in [ 17 , 32 ] do not implement activity recognition.…”

Using BlazePose on Spatial Temporal Graph Convolutional Networks for Action Recognition

“…Recently, the ST-GCN model usage has increased considerably. From stroke type recognition in tennis [28], nurse activity recognition in hospitals [29], stock price prediction [30], fall detection [31] and action recognition systems [32]. Unfortunately, this architecture presents some disadvantages.…”

BERT for Activity Recognition Using Sequences of Skeleton Features and Data Augmentation with GAN

“…In other recent cases, instead of using RGB images directly, they are used to extract poses represented by a set of body joints and their interconnection (i.e., human skeleton representation) [ 5 , 7 , 8 , 9 , 10 ], which are then used as features for further analysis. Such skeleton representations have been found powerful to differentiate between different types of activities such as walking, sitting, jumping, running, falling down, etc.…”