University of Twente Student Theses
Resource-Efficient Deep Learning for Mobile Activity Recognition on Edge Devices
Lu, Yixiang (2025) Resource-Efficient Deep Learning for Mobile Activity Recognition on Edge Devices.
This is the latest version of this item.
|
PDF
3MB |
| Abstract: | This thesis investigates methods to improve the efficiency and performance of deep-learning for sensor-based Human Activity Recognition on limited-resource edge devices, using the DeepConvLSTM as a case study. The primary objective is to reduce model size and inference time while maintaining high accuracy. To achieve this, existing techniques such as quantization, pruning, knowledge distillation, and attention mechanisms were evaluated for their effectiveness and compatibility when combined. The optimized models, integrating these techniques, were successfully deployed on the Arduino Nano 33 BLE Sense. This deployment demonstrates the feasibility of running DeepConvLSTM models enhanced with knowledge distillation and attention mechanisms on tiny TensorFlow Lite for Microcontrollers edge devices. To further improve efficiency in sensor-based Human Activity Recognition, this thesis introduced event-driven inference, inspired by Neural Architecture Search and Dynamic Routing Networks. The event-driven model dynamically selects the most suitable path for each input sample, significantly improving efficiency while maintaining high accuracy. By combining event-driven inference with existing optimization techniques, the proposed models significantly reduce inference time while achieving slightly higher accuracy than the original models. They only exhibit 7.85\% to 35\% of the inference time comparing with the original model, and the size is also largely reduced. These results underscore the effectiveness of event-driven inference in Human Activity Recognition tasks, showcasing its potential to greatly enhance efficiency and performance on resource-constrained edge devices. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Embedded Systems MSc (60331) |
| Link to this item: | https://purl.utwente.nl/essays/105044 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page