DEVELOPMENT OF PERCEPTION AND MOTION PLANNER SYSTEM IN MULTI-LANE HIGHWAY SCENARIO

• Main Author: Pratama, Satria
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67041
• Institution: Institut Teknologi Bandung

Safety is the most important factor to be taken into consideration while driving, autonomous cars included. There are crashes that happens due to environment recognition failure in an autonomous system. Recent studies are targeted to improve the system so that autonomous car be fully autonomous. Autonomous vehicle research done by Teknik Fisika Institut Teknologi Bandung (TF ITB) currently only developed a driver assists system. This final project is proposed to increase the autonomous level of the TF ITB autonomous golf cart to make decision in certain scenarios. The improvement is done by implementing the current control system with a local planner that is capable to response multiple dynamic obstacles in multi-lane highway and generate real-time waypoints. Hence, the system is useful in increasing passenger safety because the system is designed to respond to its environment. Motion planner is a component in an autonomous vehicle that updates the controller with local waypoints by using an environmental perception feedback. The proposed local planner is designed by integrating behavior tree frame work and a conformal spatiotemporal lattice planner with an object detector that uses the YOLO architecture, and an object matcher that uses the DeepSORT architecture. Data acquisition for the object detector is executed by utilizing a depth camera and implemented in the TF ITB autonomous golf cart. The integration between the controller, sensors, and actuators is achieved using Robot Operating System (ROS) as the middleware. Experiment on autonomous golfcart shows that the development of perception system using YOLO and DeepSORT algorithms, as well as Behaviour Tree algorithm based motion planner, could be implemented into autonomous vehicles. Dynamic object tracking using YOLOv5 medium + DeepSORT is accurate enough to be implemented into autonomous vehicle, having mean absolute position error of 6.9cm for experiments between -2.5 to 2.5 meters in ???????? axis and 6.1cm for experiments between 5.0 to 10.0 meters in ???????? axis.