Automatic Vehicle Trajectory Extraction from Aerial Video Data

In this paper we present a complex solution to automatic vehicle trajectory extraction from aerial video data, providing a basis for a cost-effective and flexible way to gather detailed vehicle trajectory data in traffic scenes. The video sequences are captured using an action camera mounted on a UAV flying above the traffic scene and processed off-line. The system utilizes video stabilisation algorithm and geo-registration based on RANSAC guided transformation estimation of ORB image feature sets. Vehicles are detected in scene using AdaBoost classifier constructed of Multi-Scale Block Local Binary Patterns features. The vehicle tracking is carried out by multi-target tracker based upon set of intra-independent Bayesian bootstrap particle filters specialized to deal with environmental occlusion, multi-target overlap, low resolution and feature salinity of targets and their appearance changes. The performance of the presented system was evaluated against hand-annotated video sequences captured in distinct traffic scenes. The analysis show promising results with average target miss ratio of 22.5% while keeping incorrect tracking ratio down to 20.4%.