Accurate Parking Scene Reconstruction using High-Resolution Millimeter-Wave Radar

A millimeter-wave radar has major advantages in robustness under adverse weather and illumination conditions. However, it has some concerns regarding signal noise and resolution. They make it difficult for the radar to precisely recognize the driving environments. We are challenging to reconstruct parking scenes requiring precise obstacle shape information. We have developed the 24-layer Convolutional Neural Network (CNN) originally designed from scratch using the accumulated radar reflection map for the reconstruction. 676 radar reflection maps with the ground truth for parking cars, curbs and fences were generated as the dataset by the 79 GHz UWB radar installed on our experimental vehicle. The object shapes were reconstructed by our network under various conditions compared with conventional CNNs, SegNet and U-net. Our CNN achieved a high estimation accuracy of over 97% under all conditions. Maximum outline errors of the parking cars in adjacent area were within 7.3 cm. In contrast, other CNNs degraded to less than 94 % and the outline errors increased to over 9.7 cm. Most of them had a considerable collapse of the reconstructed shape.