VEHICLE DETECTION USING MACHINE LEARNING APPROACH

• Main Author: Agung Prihatmaja, Pratamamia
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39347
• Institution: Institut Teknologi Bandung

Vehicle detection is one of the domains of research that has received much attention in the field of image processing and computer vision. This application will detect vehicles for the industry. Some examples of applications that utilize vehicle detection are traffic monitoring, automatic ticketing, and self-driving car. Previous research on general object detection has been done and produced an object detection architecture named YOLO (You Only Look Once). This architecture has a high detection speed with a fairly good degree of accuracy. However, to detect vehicles from various angles of shooting on Indonesian highways, this model is still not able to work optimally. This study tries to adapt the YOLO architecture to be used as a vehicle detection model from various angles of shooting on Indonesian highways. There are three backbones that are tried to be applied to this architecture, namely Tiny YOLO, Darknet-53, and MobileNets. All three have their respective advantages, both in terms of detection speed and accuracy. The training data used comes from the COCO dataset which has been filtered to get vehicle data only. The test data used comes from Indonesian highway data which are annotated manually. The test results show that the model with YOLOv3 architecture that adapts MobileNets backbone have the highest level of accuracy, i.e. with mean-average precision (mAP) reaching 67.77%. This shows a pretty good increase, because the YOLOv3 model that has been developed previously has a mAP of 53.25%. This study also tried to develop several methods for the post-detection process. The postprocessing method developed seeks to provide a heuristic aspect of human knowledge in the model built. The post-processing method in the form of determining the detection area and setting the bounding box area can increase mAP to 68.56%.