Human Activity Recognition (HAR) is the keystone for healthcare, smart homes, and mobile computing; nevertheless, the process of putting deep learning models right on edge devices has always been a hard nut to crack because of the high computational requirements. This research article is a perfect example of a work that solves what we might call the paradox of real-time and efficient HAR with accuracy kept intact. In this paper, this is achieved by creating a lightweight convolutional neural network (CNN) and using TensorFlow Lite to perform the quantization and related optimizations on the UCI HAR dataset.

The design reaches 93.5% of the accuracy before quantization. After the quantization, there is a loss of less than 1% in accuracy. The dimension of the model decreases from ~4.2 MB to ~0.9 MB, and the inference latency is almost half, thus making it IoT as well as mobile devices user-friendly. The findings validate the possibility of employing compact CNNs as they can strike a balance between the accuracy of the solution and computational efficiency, thus making it possible to perform HAR on platforms with limited resources.

This is a project that merges the possibility of achieving the highest accuracy for HAR and the feasibility of deploying it in edge devices. The research works that lie ahead include multimodal HAR, lightweight transformer architectures, and real-world streaming applications.