Master's Thesis : Imaging radar signal processing for human positioning and behavior classification

Abstract

This master’s thesis was conducted in conjunction with an internship at BEA, a sensor manufacturing company headquartered in Liège, Belgium. The BEA group designs and produces opening and safety sensing solutions for pedestrian and industrial doors. Nowadays, indoor space monitoring is a key aspect to improve safety and comfort for users. Surveillance of public spaces and useful data acquisition are subjects of current concern. Yet such practices do not always comply with data protection and respect for privacy. Radar signals, however, do not involve identity revealing images. Moreover, machine learning techniques are well suited to radar imaging data exploitation such that the same system can be used for 3D target imaging and human behavior classification. The object of this work is to design a human detection and positioning system capable of classifying between different attitudes and positions. This task is performed by using a Frequency-Modulated Continuous-Wave (FMCW) imaging radar module, with integrated antennas on package, manufactured by Texas Instruments. The device is installed indoors, as an overhead sensor. The radar is operated in Multiple-Input Multiple-Output (MIMO) mode across the 60 GHz to 64 GHz frequency band and its performance is evaluated. This work describes the determination of suitable operating parameters for the task. A signal processing scheme was designed, from raw signal acquisition to 3D re-projection, positioning and imaging of detected targets, with the addition of speed and power information. After some data acquisition and labeling campaigns, classification models and a modified neural network were trained to distinguish between several predetermined human activities, based on the spatial, velocity and reflected power information. The system works at different installation heights and effectively detects people in a circular area of 7m in diameter on the ground. It presents detection rates of 98% for humans in motion and of 50% to 60% for resting postures. The best class weighted accuracy score for the behavior classification task is 91% and was obtained by adapting a specific neural network architecture to five-dimensional data. The results suggest that the radar module is suited to the task and fulfills its detection, positioning, imaging and classification roles. One aspect that deserves to be improved is the detection of micro movements of static people. Several radar modules may be paired in order to increase detection area, and target tracking is feasible. The system is expected to generalize well to different setups. Exploiting temporal information from sequential records may improve classification scores.