OBJECT DETECTION AND TRACKING FOR MIXED AND DENSE TRAFFIC IN AUTONOMOUS VEHICLE

• Main Author: Satrio Wicaksono, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49325
• Institution: Institut Teknologi Bandung

Autonomous vehicle is basically depent on their environment information, thus perception system is needed for detecting object location and tracking object movement around the vehicle. But in reality, every country in the world have different traffic situation. For example, developing country such as Indonesia have more dense and mixed traffic. Hence, object detector and tracker are needed to be adjusted to work properly in those traffic condition. In this thesis, autonomous vehicle’s perception system is conducted by combining two stages that consist of object detection and object tracking. Moreover, to encounter the difference of traffic condition in Indonesia, new dataset that captured on real traffic condition in Bandung city is introduced. Object detection is utilized for locating vehicle and pedestrian on the video frame and classifying the objects. Object detection can be approached with different methods, one of the popular methods is deep learning. YOLOv3 is object detection method that performed really well in both speed and precision. In this thesis, the original YOLOv3 will be compared with YOLO-SPP which is original YOLOv3 with additional SPP layer. Furthermore, in order to detect objects that are in accordance with traffic conditions in Indonesia, the model must be retrained. Transfer learning is the best training method to obtain optimal training results. After the training and testing process have been conducted with the dataset in Indonesia, YOLO-SPP which was trained with transfer learning achieve precision value of 95.49% mAP compared to the YOLO-SPP model training from the beginning which received 92.39% mAP precision. In addition, YOLO-SPP also has better precision value than YOLOv3 with transfer learning which only obtain 93.18% mAP. In terms of object tracking, SORT algorithm is used in the autonomous car perception system. The test results of tracking with the SORT algorithm are very dependent on the object detection model used. With the YOLO-SPP and SORT models, the average MOTA was obtained at 87.73% and the average MOTP at 95.03%.