• Main Author: Ardellia Arfian, Karina
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73535
• Institution: Institut Teknologi Bandung

<p align="justify">Today, autonomous vehicles continue to grow rapidly to fulfill the development of future transportation. The innovation comes from the fact that 94% of the reasons for traffic accidents are caused by human factors, with detailed data; 41% due to recognition failure and 33% due to failure in decision making. In addition, it is also supported by the president's policy to implement autonomous vehicles as transportation in the new capital city. The number of scenarios that occur on the road and the high demand of the industry have made several companies start researching and developing autonomous cars such as Tesla, Waymo, Motional and others. These companies are trying to build Advanced Driver Assistance Systems (ADAS) technology on vehicles that are able to analyze the environment around the road just like humans. In this research, an ADAS system is developed that able to make decisions based on the position and distance of the vehicle in front of the test vehicle as well as the type of lane detected. To detect and track objects, the Yolov5 and DeepSORT methods are used, while to detect lanes, the instance segmentation method with ERFNet backbone type is used. Two different types of training datasets were used, one each for training object detection and lane detection. To train the object detection method, the COCO dataset is used while the lane detection method is trained by the TUSimple dataset. Furthermore, for the classification of the previously detected lanes, LCNet deep learning is used to classify their types based on two categories: solid and dashed lines. From the position data, the distance of the preceding vehicle relative to the test vehicle, as well as the lane type, a decision-making process is defined for the vehicle to decide whether the vehicle can change lanes to the left or right lane and also provide a collision warning. The test results of the developed algorithms using both datasets and real data provide accurate detection results. For object detection using Yolov5 and DeepSORT methods can produce accuracy up to 90%, while lane detection using ERFNet instance segmentation approach and lane classification using LCNet produce accuracy of about 96% and 74%, respectively. Experiments were conducted by recording data using camera sensors on several Bandung roads with variations of straight and turning roads. The proposed decision algorithm has also been able to provide visualization for lane change warning and collision warning.