Real-Time Multi-obstacle Detection and Tracking Using a Vision Sensor for Autonomous Vehicle

Abstract

In this paper, an effective approach for real-time multi-obstacle detection and tracking in the navigation module is discussed.  To calculate a feasible path for an autonomous ground vehicle (AGV) from the start position to goal position, efficient Dstar lite global planner is added and adhered to ROS nav$\_$core package. Later, the clustering of points based on distance from a laser scanner data is carried out to perform multi-obstacle detection and followed by tracking. Then, the clusters are categorised as static and dynamic obstacles from their location, orientation, speed and size of an individual cluster. Using this approach, dynamic obstacles’ paths are estimated from their respective past positions. To predict the dynamic obstacle for the next five-time steps, linear extrapolation and line fitting are employed. The estimated obstacles’ path data are published as a PointCloud ROS message, then it is subscribed by the costmap node of the ROS navigation package. The costmap automatically updates the obstacle map layer and rebuilds the 2D occupancy grid map with new information about obstacles. Then, the path planner replans the path using updated costmap to avoid obstacles in the dynamic environment. Finally, real-time experiments are conducted to validate the efficacy of this intelligent system.