Pedestrian detection and tracking for mobility on demand

Abstract

This paper presents a pedestrian detection and tracking system to be used aboard mobility on demand systems. Mobility on demand is a transportation paradigm in which a fleet of vehicles is shared among a community, with rides provided upon request. The proposed system is capable of robustly gathering pedestrian paths in space using 2D LiDAR and monocular cameras mounted onboard a moving vehicle. These gathered pedestrian paths can later be used to infer network traffic to learn to anticipate the location of ride requests throughout a day. This allows mobility on demand systems to more efficiently utilize resources, saving money and time while providing a more favorable experience for customers. The onboard LiDAR is used to cluster and track objects through space using the Dynamic Means algorithm. Pedestrian detection is performed on images from the mounted cameras by extracting a combination of histogram of oriented gradients and LUV color channel features which are then classified by a set of learned decision trees. Temporal information is leveraged to achieve higher detection quality by accruing classification votes. Both a standard fusion technique and a novel extrinsic calibration error-resistant fusion method are tested to fuse camera and LiDAR information for pedestrian path collection. The novel error-resistant fusion system is shown to outperform standard fusion techniques under both normal conditions and when synthetic extrinsic calibration noise is added. System robustness and quality is demonstrated by experiments carried out in real world environments, including the target environment, a university campus.