TinyML based Deep Learning Model for Activity Detection

Abstract

Our physical and emotional well-being are directly impacted by our body positions. In addition to promoting a confident, upright image, maintaining good body posture during various activities also ensures that our musculoskeletal system is properly aligned. On the other side, bad posture can result in a number of musculoskeletal conditions, discomfort, and reduced productivity. Accurate systems that can detect posture in real time, activity detection, are required due to the rising use of wearable technology and the growing interest in health and fitness tracking. The goal of this project is to create a TinyML model for wearable activity detection that will allow users to assess their posture and make necessary corrections in order to improve their health and general well-being. The project intends to contribute to the creation of useful posture detection technologies that can be quickly implemented on wearable devices for widespread usage by leveraging machine learning algorithms and wearable sensor data. For reliable posture categorization, the model architecture combines deep neural networks (DNN) and LSTM layers. With the development and implementation of the TinyML model, a significant decrease in the model's power consumption, memory, and latency was achieved without any compromise in the accuracy. This work can be used in the fields of health, wellness, rehabilitation, corporate life, sports and fitness to keep track of calories burned, activity duration, distance traveled, posture analysis, and real-time tracking.