SenseBody: A Multi-agent Human Activity Recognition System Using Wearable Devices

The exploding of the Internet of Things(IoT) has brought in an enormous number of smart devices into humans' everyday lives. People start to carry more devices on the body to monitor vital signals or increase their lifestyle convenience, for instance, wearing an Apple Watch or an Airpods. These new devices bring new opportunities for accurate body gesture recognition. However, traditional methods that rely on a single device cannot precisely recognize more complex gestures. Cooperating between multiple devices to achieve better accuracy while minimizing computation and communication costs is still a challenge for existing methods. Recognizing complex gestures with various devices is still kind of left blank.
This thesis presents SenseBody, a system that utilizes multiple wearable sensors and devices in the different body parts to enable gesture recognition. Our system utilizes machine learning techniques focusing on its easy pairing and energy preserving techniques. We use the current state-of-art machine-learning methods to classify the current dataset and collaborate different body part sensor data to make better classifications. We built a prototype based on COTs on-body wearable devices and evaluated the system's accuracy and energy performance. Our results have shown that by applying the proposed multi-agent framework, we can achieve an average of 98.3% of recognition accuracy on nine everyday activities and achieve more than 98.9% auto-pairing accuracy. From an energy-saving perspective, our result also shows that our system can save more than 90% of the energy consumed in data transmission over traditional methods while only sacrificing 3% of accuracy. Our evaluation has demonstrated great potential for future human activity recognition systems.