Evaluation of Human Activity Recognition and Segmentation Based on the DETR Model

In the field of Human Activity Recognition (HAR), segmenting sensor data streams is considered a crucial preprocessing task, typically relying on fixed-size windows. However, this approach presents two main challenges: the multi-class window problem caused by multiple activities within a fixed-size window and the fluctuation of prediction results due to noisy data and over-segmentation. The introduction of the DETR (Detection Transformer) model eliminates the need for sliding windows, simplifies the segmentation and recognition process, and masks the significant drawbacks of using sliding windows, potentially further enhancing recognition performance. This study aims to explore the feasibility of the DETR model in handling time-series data for activity recognition and segmentation tasks. By integrating the DETR model into the HAR domain, we propose a novel method for recognizing and segmenting activities in continuous time series. This method leverages the end-to-end nature of the DETR model, enabling simultaneous activity detection and segmentation, thereby improving algorithm efficiency. In the experimental section, we conducted extensive tests on several publicly available HAR datasets to validate the effectiveness of the proposed method. These experimental results fully demonstrate the potential of the DETR model in human activity recognition and segmentation tasks.