Action recognition using microwave radar sensors
Project Team: Doğa NALCI, Prof.Dr. Yusuf Sinan Akgül
Action recognition using microwave radar sensors
Prediction of human movements is an important issue that is needed in many fields and that researchers are working on. Although there are solutions using camera, wearable technologies and sensor technologies in the literature, these solutions are highly dependent on the position of sensors and intensity of the light beams reflected from the objects in the environment. Camera sensors cause serious privacy violations. Wearable sensor technologies, on the other hand, have limitations such as not being able to provide sufficient information about the entire skeletal structure. Therefore, these methods are not practical and not preferred. In this thesis, human movements were predicted using Frequency Modulated Continuous Waves (FMCW) millimeter wavelength (MMWAVE) radar. Instead of the point cloud obtained from the radar, the raw data formed in the radar receiver antennas are focused. The raw data created by the human movements in a certain time period on the Mmwave radar receiver antennas were recorded and converted into a multichannel and multi-dimensional form using the visualization technique we call "VIDAR". This data was first used in training our Support Vector Machine model and then used in training the CNN network we finetune. A real-time test environment has been created to test the trained model with data it has never seen before. As a result of tests conducted in this environment, it has been observed that human movements are predicted with an accuracy of over 90%.
Figure 1: Rawdar data operations
Detection of human movements is an important issue in many areas. Detection of movements such as fight and escape is very important for security systems. In automotive industry, detecting the actions of both drivers and pedestrians will prevent accidents that may result in injuries or death. Sales can be increased by making video analysis of human motion in retail stores and areas where human mobility is intense. In rehabilitation treatment applications, the movements must be monitored and controlled by an expert with feedback.
Figure 2: Rawdar Workflow