Parking Monitoring Based on FMCW Radar Imaging with Deep Transfer Learning

Abstract

In this paper, we applied the TI 77GHz mmWave Frequency Modulated Continuous Wave (FMCW) radar for the occupancy classification of parking slots with edge deep learning using radar images of both indoor (basement) and outdoor parking areas. The platform used for edge deep learning is based on the Qualcomm 865 processor. To save the required data and training time, transfer learning (based on the trained model for indoor parking areas) is employed for the classification of outdoor parking slots. A convolutional neural network (CNN) is used for the multi-class classification of FMCW radar signals. According to the performance evaluation for the multi-class classification using accuracy, around 99.99% accuracy and 0.01 categorical cross-entropy loss are achieved (in both the training and validation steps with 300 epochs) for the occupancy detection of indoor parking slots, and more than 92 % accuracy and 0.03 cross-entropy loss for that of outdoor parking slots. With the transfer learning, the data size required for the classification of outdoor parking slots is about 37% for that of the indoor one.