Pedestrian detection and target classification in urban forestry environment with FMCW automotive radar and CNN

This thesis studies the pedestrian detection in complex urban environments by using Frequency Modulated Continuous Wave (FMCW) automotive radar with a particular focus on some challenges caused by obstacles such as trees and vehicles. While some studies rely on simplified radar target models, this research introduces realistic 3D representations of forestry environment to simulate driving scenarios that represents real world conditions. A significant component of this work includes Radar Cross Section (RCS) analysis across various aspect angles to understand how wind-induced motion in tree model affects radar signal integrity. The dynamic behavior of the branches of trees under windy conditions introduces additional clutter and micro-Doppler noise which can adversely impact pedestrian detection performance. To address these issues, convolutional neural network (CNN) is employed for the extraction and classification of micro-Doppler signatures from radar spectrograms. The CNN model is designed and trained to distinguish pedestrian features from signals which are under environmental interference. The performance evaluation is performed by using metrics such as accuracy function, lossfunction and confusion matrices. Experimental results indicate that the proposed CNN based detection method provides robust classification in the presence of environmental noise and distortions. When it is compared to traditional approaches, the detection and the classification system demonstrates improved reliability and detection accuracy under realistic conditions. All in all, this study contributes to the development of pedestrian detection systems by highlighting the importance of realistic and proper scenario modeling. It is expected that the results provide valuable conception for the development of safer and more resilient perception systems in autonomous driving applications.