Optimization of mmWave radar image classification by YOLOv7 using the Pix2pix architecture

The classification of objects detected by mmWave radar has become more and more treated. In this work we tried to optimize the classification, that by cleaning the heatmap images of the fluctuations that prevent the good classification. First, we captured a dataset, it consists of more than 1510 frames taken for different objects and in different environments. Each frame is a $100\times100$ matrix, which will be represented later as a heatmap. Then we annotated the objects present on the heatmaps. We try two different methods, the first one is the training and the classification of YOLOv7 without denoising. The second one is the cleaning of the heatmaps by the Pix2Pix, then we classify, at last we compare the results. For the classification we propose a model named YOLOv7-PM, which was compared with several standard models. We notice when comparing the results, that the images cleaned by Pix2Pix are better classified than the original images. We reach a mAP_0.5 of 0.92 for our model with the cleaned images, and of 0.89 for the original images.