Modeling the impact of observation conditions on localization systems

In various applications, sensor fusion has demonstrated success as means to enhance a system performance in perceiving its environment. By combing observations of different sensors, the system is able to achieve improved sensing accuracy, and potentially, expanded sensing capabilities. However, the observation conditions in the surrounding of any multi-sensor system have a considerable impact on the performance of the system. This impact can be hard to mitigate if the observation conditions are stochastic in nature. Therefore, for any sensor fusion strategy to achieve reliable and robust performance it must possess a capability to assess the quality of the observation conditions in its surrounding, and ultimately, the quality of its decisions, as a function of these conditions. One typical application where the impact of the observation conditions can cause sever deterioration of the sensing performance is vehicle localization. It is typical in this application that location measurements obtained from multiple sensors (e.g., GPS, Vision, Inertial, etc.) are combined together to compute accurate vehicle location. However, such improved accuracy can only be attained under nominal observation conditions. Therefore, real-time awareness of the observation conditions around the vehicle position is pivotal for the multi-sensor system to achieve effective fusion performance. In this paper, a Markovian model is proposed to capture the impact of observation conditions on a sensor's localization performance and to consequently determine a reliability index with respect to the localization accuracy claimed by the sensor. The proposed model is implemented on two localization techniques: single-sensor localization and multi-sensor localization. A number of experiments are conducted to determine the different levels of localization accuracy that can be achieved by each technique under a wide range of observation conditions. The proposed reliability model is tested in a variety of real-life and simulated observation conditions scenarios. It is evident from the experimental results that the proposed model is able to estimate the reliability of location estimates produced by either one of the localization techniques. The paper discusses how such reliability model can benefit multi-sensor systems.