| Popis: |
This work focuses on the estimation of the ground-plane parameters needed to rectify and reconstruct crowded pedestrian scenes, projected into 2D by an uncalibrated, monocular camera. Deformities introduced during the imaging process affect metrics such as size, velocity and distance, which are often useful when examining the behaviour of agents within the scene. A framework is presented to reverse “perspective distortion” by calculating the “groundplane”, upon which motion within the scene occurs. Existing methods use geometric features, such as parallel lines, or objects of known size, such as the height of individuals in the scene; however these features are often unavailable in densely crowded scenes due to occlusions. By measuring only the imaged velocity of tracked features, assumed to be constant in the world, the issue of occlusion can be largely overcome. A novel framework is presented for estimation of the ground-plane and camera focal-length for scenes modelled with a single plane. The above assumption is validated against simulations, outperforming an existing technique [12] against real-world benchmark data. This framework is extended into a two-plane world and the additional challenge of determining the respective topology of the planes is introduced. Several methods for locating the intersection-line between the two planes are evaluated on simulations, with the effect of variation in velocity and the height of tracked features on reconstruction accuracy being investigated, with the results indicating this technique is suitable in real-world conditions. This framework is generalised, removing the need for prior knowledge of the number of planes. The problem is reformulated as a linear-series of planes, each connected by a single hinge, allowing the calculation of a single rotation for each new plane. Again, results are shown against simulations on scenes of varying complexity, as well as realworld datasets validating the success of this method given realistic variations in velocity. |
