Image Based Place Recognition and Lidar Validation for Vehicle Localization

In this paper, we propose a system for vehicle localization that combines two sensors: a camera and a lidar. An image based place recognition approach is used to determine the vehicle localization when the vehicle revisited a previously visited location. Unlike systems that only rely on visual appearance recognition for localization, we also integrate lidar measurements information in order to validate the vision based place recognition results. Effectively, false positives recognition can be detected and rejected by checking the coherency of the image based recognition results with the results of lidar measurements matching with ICP (iterative closest point) algorithm. In case of false image based recognized places, vehicle position can be computed using only lidar based ICP method. The vehicle position is effectively estimated using the last known position and the transformation between the corresponding lidar measurement and the current one obtained by applying ICP. By employing the camera and lidar sensors, the deficiencies of each individual sensor can be overcome. Experiments were conducted in two different surrounding areas. The obtained results show that the proposed method permit to avoid the well-known long-term accumulated error of deadreckoning localization and lidar data can help to reject false positives of place recognition.