Vehicle detection and classification from mobile LIDAR sensors

This empirical proof-of-concept study uses LIDAR sensors mounted on a moving instrumented probe vehicle to segment passing vehicles from the LIDAR returns, cluster the returns into distinct vehicles, and then classify the clusters into specific vehicle types. This study is among the first to use LIDAR sensors mounted on a moving vehicle to classify the ambient vehicles into categories suitable for roadway traffic monitoring applications. The extracted vehicle clusters are classified using two methods: a low-resolution height-and-length based methodology (HLM) and a higher resolution scheme that extracts vehicle silhouettes for a shape-based classification methodology (SBCM). The SBCM standardizes each vehicle point cloud into a silhouette that captures the vehicle’s profile. The silhouettes are compared to prototypes for each class that were derived from representative vehicles in the development set. In the 7000 vehicle validation data set HLM had 99.0% accuracy and SBCM had 93.8–99.2% accuracy (depending on the number of classes). This process runs faster than real time, employing techniques to simplify the processing without sacrificing performance. While this work used paired 2D LIDAR sensors, we rectify the 2D measurements over time to create a 3D point cloud, as such, the general methodology should be transferrable to 3D LIDAR sensors. It is envisioned that this new means of collecting classification data can extend surveillance to currently unmonitored links, supplementing the existing fixed location classification sites. Furthermore, each time a mobile sensor passes a fixed location station, the two systems can be used to automatically cross-validate one another.