A Traffic Sign Classification Method Using LiDAR Corrected Intensity and Geometric Feature

As an important perception sensor for autonomous vehicles (AVs), light detection and ranging (LiDAR) provides 3D-spatial and 1D-intensity information. To boost the ability of traffic sign classification (TSC) using LiDAR, a novel classification method combining corrected intensity and geometric feature was proposed to identify traffic sign patterns. An unequal-interval-division (UID)-based intensity frequency histogram (IFH) was advanced to form high-quality input features, thus facilitating the optimization of the backpropagation neutral network (BPNN) classifier for better performance. A series of experiments was conducted, including ablation study and parametric investigations involving point density, sign patterns, and instrument types. Results showed that the combination of geometric and corrected intensity (UID-IFH) feature enhanced the classification performance significantly, with the indicator F1 score achieving 0.812–0.921 at the points density of 0.03–0.91 pt/cm $^{\mathbf {{2}}}$ . Compared to the state-of-the-art commercial in-vehicle LiDAR with unreliable intensity, the F1 score obtained from the high-stable intensity LiDAR has an obvious improvement. The proposed method is expected to obtain better performance with the advancements in LiDAR technology (e.g., high-density imaging in compact size and accurate intensity of low-cost laser), to serve as an effective approach for decision-making in AVs.