Embeded Signal Processing Lab

Human Activity Recognition (HAR)

- Human Activity Recognition (HAR) is a technology to identify specific movements or actions of an individual or group using various sensors (image, inertial sensor, radar).

- Radar-based HAR technology performance using deep learning has only proven its feasibility through experiments in a limited environment. 
   In indoor environments, recognition performance deteriorates by 70-80% due to changes in geometric structure and appearance of human information.
- We aim to improve behavior recognition accuracy and perform robustly despite changes in the indoor environment and installation location. We propose a radar-based HAR that combines Range-Time-Doppler (RTD) maps and Range Distributed (RD)-CNN.

IPS is a technology that estimates the location of a device in a cyber-physical system

- IPS is a technology that estimates the location of a device in a cyber-physical system and is operated based on general-purpose Wi-Fi.
- However, due to the diversity of the positioning service area, the radio map is not standardized, so Fingerprint is not universal and does not guarantee stability in calculation amount and accuracy.
- We aim to propose a general-purpose fingerprint system by dividing the service area into optimal regions and standardizing the radio map.

Inertial Odometry is a technology that estimates the amount of position change using an IMU

- Inertial Odometry is a technology that estimates the amount of position change using an IMU and is being studied separately into 3-DOF IO (2D odometry) and 6-DOF IO (3D odometry).

- 6-DOF IO requires more axis data to be estimated than 3-DOF IO, it undergoes an integration process when creating a trajectory by estimating the amount of change in position and posture, resulting in a drift error.

- We aim to improve posture estimation accuracy and trajectory estimation accuracy. We propose the method that improves the trajectory estimation accuracy by improving the posture estimation accuracy by additionally inputting gravitational acceleration and geomagnetism and reduce the drift error.