Augmented Map Based Traffic Density Estimation for Robot Navigation

Most of the work done on robot navigation is focused on map building, localization and obstacle avoidance. In this paper, we particularly focus on intelligent robot navigation based on scene traffic. In this course we initially generate augmented maps using scene traffic flow captured at discrete times of same days, different days and various weather conditions. For facilitating the same, we manually navigate a mobile robot(X80SV) for capturing data(image and sonar map data) across a campus; segmented into zones. Zone specific point cloud maps are then created by merging scene object detection results with sonar map. A Gaussian model is further used for estimating trend for traffic density from the point clouds. This information is later utilized for navigational purposes, to detect the minimum crowded path from a source to destination. Object detection and classification is done using a fine tuned AlexNet. For labelling the training set used for fine tuning, we perform foreground extraction, transfer learning and SVM classification of the scenes. 16 classes of objects including background is used. During realtime navigation, we also perform object detection and classification so that the current traffic trend can be matched with the augmented maps.