YODA

The problem of Human Activity Detection (HAD) is to detect every occurrence of a target activity within a sensor time series. Unlike Human Activity Recognition (HAR) algorithms, which map a segment of a time series onto an activity label, HAD algorithms provide labels of activity occurrences together with the start and end time of each occurrence. HAR models have demonstrated high performance in controlled laboratory datasets, and their success highlights the importance of leveraging pretrained recognition models to enhance HAD systems. In this research, we introduce YODA (You Only Detect Activities), a unified framework that formulates HAD as a temporal object detection problem inspired by computer vision. We present the problem of activity detection and explain how the YODA algorithm processes sensor time series to detect target activities. We evaluate YODA on existing public datasets. Additionally, we further evaluate performance for a new dataset that we collect for individuals labeling activities in the wild. To improve performance, we further consider improvements to YODA that build upon the strengths of existing HAR backbones. Experimental evaluations show that YODA effectively detects locomotion and static activities, while challenges remain for brief transitional actions. By integrating pretrained HAR knowledge into detection tasks, this work bridges the gap between recognition and detection. YODA offers a novel human activity detection method for real-world human activity analysis using wearable sensors.