Perception Based Navigation for Autonomous Ground Vehicles

Autonomous vehicles rely on their sensors specializing in specific tasks to understand their surroundings. Camera is perhaps one of the most versatile of them. Frames from a camera, when processed, form the basis for object detection in the surroundings, which is an embodiment of perception. The major tasks in perception are object tracking and localization. Neural Networks, especially Convolutional Neural Networks (CNN), exceed expectations in the tasks of detection and classification. This paper adds to the functionality of CNN by exploring the aggregation of additional knowledge from images, consisting of segmentation and localization information, as an attempt to predict safe driving zone for robust navigation. This aggregated knowledge can be visualized as a meta representation of the surroundings for the navigation algorithm.