LIDAR Based Vehicle Classification

Vehicle classification data are used for numerous transportation applications. Most of the classification data come from permanent in-pavement sensors or temporary sensors mounted on the pavement. Moving out of the right-of-way, this study develops a LIDAR (light detection and ranging) based classification system with the sensors mounted in a side-fire configuration next to the road. The first step is to distinguish between vehicle returns and non-vehicle returns, and then cluster the vehicle returns into individual vehicles. The algorithm examines each vehicle cluster to check if there is any evidence of partial occlusion from another vehicle. Several measurements are taken from each non-occluded cluster to classify the vehicle into one of six classes: motorcycle, passenger vehicle, passenger vehicle pulling a trailer, single-unit truck, single-unit truck pulling a trailer, and multi-unit truck. The algorithm was evaluated at six different locations under various traffic conditions. Compared to concurrent video ground truth data for over 27,000 vehicles on a per-vehicle basis, 11% of the vehicles are suspected of being partially occluded. The algorithm correctly classified over 99.5% of the remaining, non-occluded vehicles. This research also uncovered emerging challenges that likely apply to most classification systems: differentiating commuter cars from motorcycles. Occlusions are inevitable in this proof of concept study since the LIDAR sensors were mounted roughly 6 ft above the road, well below the tops of many vehicles. Ultimately we envision using a combination of a higher vantage point (in future work), and shape information (begun herein) to greatly reduce the impacts of occlusions.