Visual localization in challenging environments

Abstract

Visual localization, the method of self-localization based on camera images, has established as an additional, GNSS-free technology that is investigated in increasingly real and challenging applications. Particularly demanding is the self-localization of first responders in unstructured and unknown environments, for which visual localization can substantially contribute to increase the situational awareness and safety of first responders. Challenges arise from the operation under adverse conditions on computationally restricted platforms in the presence of dynamic objects. Current solutions are quickly pushed to their limits and the development of more robust approaches is of high demand. This thesis investigates the application of visual localization in dynamic, adverse environments to identify challenges and accordingly to increase the robustness, on the example of a dedicated visual-inertial navigation system. The methodical contributions of this work relate to the introduction of semantic understanding, improvements in error propagation and the development of a digital twin. The geometric visual odometry component is extended to a hybrid approach that includes a deep neural network for semantic segmentation to ignore distracting image areas of certain object classes. A Sensor-AI approach complements this method by directly training the network to segment image areas that are critical for the considered visual odometry system. Another improvement results from analyses and modifications of the existing error propagation in visual odometry. Furthermore, a digital twin is presented that closely replicates geometric and radiometric properties of the real sensor system in simulation in order to multiply experimental possibilities. The experiments are based on datasets from inspections that are used to motivate three first responder scenarios, namely indoor rescue, flood disaster and wildfire. The datasets were recorded in corridor, mall, coast, river and fumarole environments and aim to analyze the influence of the dynamic elements person, water and smoke. Each investigation starts with extensive in-depth analyses in simulation based on created synthetic video clones of the respective dynamic environments. Specifically, a combined sensitivity analysis allows to jointly consider environment, system design, sensor property and calibration error parameters to account for adverse conditions. All investigations are verified with experiments based on the real system. The results show the susceptibility of geometric approaches to dynamic objects in challenging scenarios. The introduction of the segmentation aid within the hybrid system contributes well in terms of robustness by preventing significant failures, but understandably it cannot compensate for a lack of visible static backgrounds. As a consequence, future visual localization systems require both the ability of semantic understanding and its integration into a complementary multi-sensor system.