A robust visual SLAM system for low-texture and semi-static environments

Precisely estimating a vehicle's pose in a prior map is a fundamental capability for autonomous driving. This task, however, remains challenging in low-texture and semi-static environments where the number of available features is limited and their attached objects are faced with the risk of positional change over time. To cope with the adverse conditions, we propose an innovative visual SLAM method that can correct the position of biased objects in the prior map besides estimating the accurate camera pose. Our approach is based upon the structure of independent object management to complete localization and mapping. Each object in the current frame is represented by a group of organized raster points. Owing to the adaptive pixel-based object rasterization, these raster points are more dependable to use in the visual SLAM pipeline. Assisted by the factor graph optimization, each object can be separately supervised although with multi-constraints from surroundings, resulting in robust tracking and map correction. With the benefit of GPU acceleration in data association, the execution time is sharply reduced. The proposed algorithm is validated in computer graphic render datasets.