Error Correction Method for Untrained Data to Estimate Accurate Parking Vehicle Shape by Millimeter-Wave Radar with Deep Learning

Abstract

We research parking scene reconstruction by a deep neural network (DNN) using a millimeter-wave radar. High accuracy can be achieved for the training dataset; however, it is degraded on untrained data. It is an unavoidable challenge of generalizability in machine learning. To solve this issue, we propose the method to utilize the model knowledge and estimate the reliability of the reconstruction results. The vehicle shape model with the Bezier curve is applied, and the shape parameters are estimated by DNN. The reliability is estimated by a generative model, Variational Auto Encoder (VAE) trained by the training dataset. When the reliability is low, the estimated model parameter is corrected with a stable deductive model. The correction effectiveness was confirmed for the experimental data measured in actual parking scenes.