Parking Vehicle Shape Estimation Using Millimeter-Wave Radar Applying Semi-Supervised Learning

Deep neural networks can achieve high recognition accuracy only with enough amount of data. However, it is sometimes difficult to prepare sufficient amounts of datasets due to the high cost of annotation. In this case, semi-supervised learning (SSL) can provide an effective means of leveraging data with no ground truth to improve recognition accuracy. Various SSL methods have been proposed. Most of them are working on classification problems that are easy to generate pseudo-labels, and no research has been done to estimate object shapes using millimeter-wave radar. In this paper, we propose a novel SSL method to realize it with the convolutional neural network (CNN). The algorithm combines three ideas. The first one is to utilize the similarity with trained input patterns to generate pseudo-ground truth. The similarity is evaluated by selfsupervised learning applying Variational Autoencoder (VAE). The second one is to utilize the consistency of estimation results with the model knowledge of ground truth. The third one applies consistency regularization by adding multiple perturbations to the input radar patterns. This combination algorithm can synergistically improve the recognition accuracy by iteratively training the CNN while adding pseudo-ground truths. As the result of the experiment using the dataset measured in the actual parking lots of city areas, the accuracy of parking vehicle shape estimation has been improved from 85.66% to 92.09% when trained with only 5% of the dataset. With 20% supervised training data, the accuracy was equal to or better than the marginal one of the conventional supervised learning (SL) trained with 80% data.