MIMO-FMCW Radar-Based Parking Monitoring Application With a Modified Convolutional Neural Network With Spatial Priors

Radar imaging is a competitive option for smart city applications over optical approaches, as it raises no privacy concerns. The inherent difficulty of interpreting radar signals can be overcome using deep learning techniques to leverage the capabilities of monitoring sensors with a minimum of human intervention. In this paper, we use a modified convolutional neural network (CNN) for classifying radar images in order to detect vacant parking spaces with a 77-GHz imaging radar. Although training CNNs for radar-image classification is challenging due to poor generalization performance caused by the lack of labeled training data, the modified architecture takes into account the properties of the radar image in order to introduce prior information into the model and improve performance. A MIMO-FMCW radar is utilized to render a slant-range image of a parking scenario, and the image patches corresponding to each parking location are classified independently in the CNN. Since the radiation pattern of a MIMO array varies as a function of the scanning angle, the corresponding spatial coordinate of each patch is included as an additional feature in the upper layers of the network. This allows the model to combine local features from each patch with global scenario information in order to learn robust features that generalize properly to new scenarios. Several models are trained end to end with data from four different parking scenarios and evaluated in a 4-fold cross-validation scheme, and performance is improved when spatial prior information is included.