Resource-Efficient Deep Learning for Real-Time Recognition of Worker Activities During Industrial Assembly

Description

This thesis focuses on developing resource-efficient deep learning techniques for real-time recognition of worker activities in industrial assembly, optimized for low-power, low-latency edge devices. A baseline 3D MobileNet model is evaluated on the MECCANO dataset, with optimization techniques like pruning, knowledge distillation, and quantization applied to enhance efficiency while maintaining accuracy.

The optimized models are tested on resource-constrained devices, such as Raspberry Pi, to ensure real-world applicability. The study also addresses dataset-specific challenges and, if time permits, explores larger datasets and advanced transformer-based models. This research aims to advance smart manufacturing by enabling efficient, real-time activity recognition for improved productivity and safety.