Object detection, segmentation, and distance estimation using YOLOv8
Akwiwu, Queen (2025)
Avaa tiedosto
Lataukset:
2025
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025052114149
https://urn.fi/URN:NBN:fi:amk-2025052114149
Tiivistelmä
The purpose of this thesis was to implement and evaluate a real-time object detection and depth estimation system using the YOLOv8 model, an Intel RealSense D435 depth camera, and a Raspberry Pi AI Hat. The goal was to improve object recognition accuracy in embedded systems by combining RGB data with depth information.
The implementation involved capturing synchronized RGB and depth streams from the RealSense camera, running YOLOv8 to detect and segment objects, and calculating distances using two methods: bounding box centroid depth and segmentation mask-based depth. The system displayed detected objects in real time, showing both visual detection outputs and their respective depth estimations.
The results demonstrated that segmentation-based depth estimation provided more accurate distance measurements compared to the bounding box method. Although deploying the model on the Hailo AI Hat presented challenges such as limited processing power and model compatibility issues, the CPU-based implementation functioned reliably.
This thesis concluded that integrating depth information with YOLOv8 segmentation significantly improves object detection performance in embedded environments. The system shows potential for applications in robotics, automation, and real-time perception tasks. Future development may focus on optimizing the model for hardware acceleration and refining segmentation-based depth calculations.
The implementation involved capturing synchronized RGB and depth streams from the RealSense camera, running YOLOv8 to detect and segment objects, and calculating distances using two methods: bounding box centroid depth and segmentation mask-based depth. The system displayed detected objects in real time, showing both visual detection outputs and their respective depth estimations.
The results demonstrated that segmentation-based depth estimation provided more accurate distance measurements compared to the bounding box method. Although deploying the model on the Hailo AI Hat presented challenges such as limited processing power and model compatibility issues, the CPU-based implementation functioned reliably.
This thesis concluded that integrating depth information with YOLOv8 segmentation significantly improves object detection performance in embedded environments. The system shows potential for applications in robotics, automation, and real-time perception tasks. Future development may focus on optimizing the model for hardware acceleration and refining segmentation-based depth calculations.