Particle Filters for Airborne Tracking and Lane-Level Map-Matching of Vehicles

Particle filters have emerged as a powerful method for solving multi-hypothesis state estimation problems, which originate from non-linear, non-Gaussian dynamic systems. Two applications of particle filters in the fields of remote sensing and vehicle navigation are addressed in this work. The first application in remote sensing is given by tracking vehicles in airborne images for traffic Monitoring and surveillance. The second application is related to vehicle ego-localization on a digital map for navigation and driving assistance systems. For airborne vehicle tracking, the low image resolution and the small frame-rate of industrial airborne camera systems impede clear discrimination of individual vehicles and precise motion prediction. Existing particle filter approaches for airborne vehicle tracking lack advanced strategies to achieve robustness against errors from low-cost systems and fail to address issues in real-world situations. Previous work on airborne vehicle tracking assumes continuous vehicle motion and clear discrimination of individual vehicles with unvarying appearance, which is not given in urban traffic scenarios. In this work, adaptive particle sampling and weighting approaches are proposed to ensure that particles explore the search space more efficiently and appearance changes can be tolerated without target losses. An update strategy for the appearance template model of each tracked vehicle further increases the rate of successfully tracked vehicles. Information about the context of multiple tracked vehicles is integrated by an online map learning approach. Experiments were conducted on low frame-rate, low- resolution image sequences taken from an airplane and a UAV of dense urban traffic scenarios with up to 151 simultaneously tracked vehicles. The results demonstrate the robustness of the developed tracker against discontinuous vehicle motion, appearance changes and difficult discrimination of similar close vehicles. For vehicle ego-localization, the standard automotive positioning sensors of series production vehicles and the abstracted representation of the road geometries in commercial digital maps complicate the computation of the correct map-matched position. Recent particle filter methods for map-matching rely on the quality of the map and the absolute positioning systems and ignore to examine the performance at ambiguous situations where correct map-matching is most critical for navigation functions. This work includes advanced models for lane-level particle sampling constrained to the digital map and particle weighting with measurements from the onboard lane detection camera system. A new medoid-shift clustering method is proposed to extract multiple hypotheses from the weighted particles in the road map space. A functional prototype of the map-matching system is implemented in a test vehicle using commercial map data and a low-cost absolute positioning system. The results of various tests including ambiguous situations like road bifurcations demonstrate the ability of the map-matcher to maintain multiple hypotheses until the situation becomes decidable. This investigation has revealed that the lane matching together with the lane topology-aware particle sampling supports the decision for the correct road segment. Overall the developed map-matcher proves its suitability for automotive applications, because of its good performance under the proposed quality criteria derived from requirements of navigation and advanced driving assistance systems. Both particle filter methods achieve robust results on experimental real-world datasets, which were obtained from low-cost sensor systems. Therefore, they prove as suitable candidates for industrial applications