A New Pavement Recognition Method of 24-GHz Radar Based on Prior Knowledge and Data-Driven

Accurate and efficient road recognition is very important for the control of mobile robots and autonomous vehicles. In this article, a new road surface recognition method based on 24-GHz millimeter-wave radar is proposed, which has better environmental adaptability compared with machine vision and absolute cost advantage compared with lidar. The core of our method is to propose a radar feature fusion method based on prior knowledge and data-driven. First, the echo signal of radar is subjected to statistical analysis, thereby confirming the distinguishability of radar signals for various road types. Then, we extract 8-D statistical features as prior knowledge features based on statistics. Second, we have designed a new representation method of radar data, which reconstructs the radar data in time series based on graphical modeling and transforms the discrete radar data into an image representation. Then, the efficient network Inception-v3 and transfer learning are used to extract data-driven features from graphical radar data. Subsequently, the feature-level fusion of prior knowledge features and data-driven features is performed to generate the feature vector that can be trained. Finally, we built the road recognition classifier based on the advanced machine learning model and used different road environments to test the effectiveness of the model. The experimental results show that our method achieves 90.6% recognition accuracy and 32-frames/s inference speed under a 24-GHz radar with a cost of only U.S.
${\$}$
16, which can be widely used in mobile robots and autonomous vehicles.