DEVELOPMENT OF LOCAL PLANNER USING DEPTH CAMERA FOR AUTONOMOUS VEHICLE CONTROL

• Main Author: Ardellia Arfian, Karina
• Format: Final Project
• Language: Indonesia
• Subjects: Teknologi
• Online Access: https://digilib.itb.ac.id/gdl/view/55289
• Institution: Institut Teknologi Bandung

Safety while driving is a factor that must be prioritized, including in autonomous vehicles. However, in reality there are still many accidents that occur due to autonomous systems that fail to respond to their environment, so recent research is aimed at improving the system and targeting the level of autonomous vehicles to be level 5. Autonomous vehicle research by Engineering Physics Study Program, Bandung Institute of Technology (TF) ITB) currently has only reached the autonomous level 2 to 3. This research is proposed to increase the autonomous level of the ITB TF autonomous golf vehicle to level 3 to 4. This is done by implementing a controller design with a local planner that is able to produce tracks in real-time. Thus, this system will be useful in increasing the safety level of drivers because the system is designed to respond to the environment. The local planner is one part of the autonomous vehicle control system which will provide updates to the track by feedback the environmental perception. The proposed local planning system is designed by integrating conformal spatiotemporal lattice algorithms and object detection using the Faster R-CNN architectural model. Data collection for object detection is carried out using a depth camera and implemented in an autonomous golf vehicle. Realization of integration between controllers, sensors, and actuators is carried out by using the Robot Operating System (ROS) as middleware. The object detection test using the Intel RealSense D435 and D455 depth cameras showed the best performance on the D455 camera with the largest absolute error of 10.29%. Therefore, object detection in the final project is carried out using a D455 camera. The local planning system was tested on a straight line without objects by providing sample frequency variations to see the vehicle's response. The results of the golf cart's optimal response were obtained at a frequency of 0.3 Hz which produced the MAE, MSE, and RMSE values ??of the golf cart's position error to the global trajectory respectively [0.40; 0.19; 0.44] and a travel time of 28.9 seconds. In addition, testing is carried out using objects on the track. The optimal response is generated using a sample frequency of 0.3 Hz, a lookahead distance of 10 m, and a weighting factor value of 0.75 with a success rate of 100%. This final project has produced a local planner that makes it possible to produce real-time trajectories based on the perceptions obtained from the depth camera. Performance evaluation for each parameter has been presented in this final project, so that it can be a reference material for local planning research on autonomous vehicles in the future.