Temporal action segmentation (TAS) is essential for identifying when actions are performed by a subject, with applications ranging from healthcare to Industry 5.0. In such contexts, the need for real-time, low-latency responses and privacy-aware data handling often requires the use of edge devices, despite their limited memory, power, and computational resources. This paper presents OLORAS, a novel TAS model designed for real-time performance on edge devices. By leveraging human pose data instead of video frames and employing linear recurrent units (LRUs), OLORAS efficiently processes long sequences while minimizing memory usage. Tested on the standard Assembly101 dataset, the model outperforms state-of-the-art TAS methods in accuracy with 10x memory footprint reduction, making it well-suited for deployment on resource-constrained devices.

Recent evolutions of recurrent neural networks (RNN) such as S4, S4D, and LRU, have shown remarkable potential for very long-range sequence modeling tasks for vision, language, and audio. They have shown a capacity to capture dependencies over tens of thousands of steps. Unlike transformers, which face significant memory consumption challenges with large context sizes, they are a promising alternative with their ability to operate effectively on embedded systems. While they have been evaluated for classification and segmentation tasks, no work in the literature has applied them in the context of human pose estimation. In this work we propose an architecture that combines such state space models (SSM) to graph attention networks (GAT) to enable their application to evaluate human action tasks on embedded systems.