Sidewalk extraction using deep learning and cost-based route optimization with mini-max objective function
With the growing diversification of modern urban transportation options, such as small-scale autonomous delivery vehicles, autonomous patrol robots, e-bikes, and e-scooters, sidewalks have gained newfound importance as critical features of High-Definition (HD) Maps. Since these emerging modes of transportation are designed to operate on sidewalks to enhance public safety, there is an urgent need for efficient and precise sidewalk annotation methods for HD maps. This is crucial for accurate representation and the development of robust path-planning algorithms for autonomous vehicles to navigate urban environments safely. The following thesis proposes a semantic segmentation-based sidewalk extraction on aerial images method using an A* path planning algorithm for sidewalk segmentation refinement. The A* path planning algorithm with and without heuristic function was then applied to the extracted and refined sidewalk annotations to generate a safe and efficient route for autonomous navigation. An objective function considering travel distance and safety level is also proposed to determine the optimal route on the sidewalk and crosswalk. The results of this work show that the proposed sidewalk extraction method can precisely and efficiently predict sidewalks from aerial images, and it is feasible to navigate throughout the city using the predicted sidewalks.