Object detection and tracking (in car cabin)

• Main Author: Huda, Md Tanvirul
• Other Authors: Yap Kim Hui
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Object detection; Tracking
• Online Access: https://hdl.handle.net/10356/177193
• Institution: Nanyang Technological University

The rate of fatalities and accidents in Singapore is increasing. Concerningly there has been a spike in the traffic accidents resulting in injuries and fatalities by 2.4% and 26%, respectively just from the year 2022 to 2023. Safety measures needed to be taken to aid in reducing such occurrence for the safety of all road users. One such measure is to utilize technology to come out with safety measures. With the rapid rise in the computer vision field there are interest to use some of the groundbreaking algorithms and models to come out with a way to make it safer for road users. To utilize these methods there needs to be an accurate model established to gather accurate data from within a car cabin. In light of making accurate detection in a car cabin this report will investigate if adding a tracker module on top of the existing object detection can improve the overall accuracy in detecting and forming bounding boxes on the target objects. The report hopes to tackle existing issue of unstable bounding boxes, false positives detection and wrong data association in different frames through this method. With the aid of 2 You Only Look Once (YOLO) models, this report has successfully concluded that adding a tracker in conjunction with object detection improves the accuracy of tracking objects across frames. This is evident by the increase in average mAP50-95 values when tracking is used (0.705 for YOLOv9 and 0.762 for YOLOv8 with tracking) compared to without tracking (0.692 for YOLOv9 and 0.743 for YOLOv8).