Improved YOLOv8 Model for a Comprehensive Approach to Object Detection and Distance Estimation
Zu Jun Khow, Yi-Fei Tan, Hezerul Abdul Karim, H.A. Abdul-Rashid
Abstract
The rapid advancements in deep learning have revolutionized the field of computer vision. However, despite the significant progress in computer vision, there remains a scarcity of research focused on utilizing this technology for distance estimation. Exploring such studies can bring immense convenience to people, especially in applications like anomaly object detection. On that account, this research proposes an improved detection model based on You Only Look Once version 8 (YOLOv8) namely YOLOv8-CAW, which is capable of both detecting target objects and accurately calculating their distances. The proposed method involves incorporating the Coordinate Attention and Wise-IoU into the YOLOv8 network, enhancing the detection accuracy. Combined with the distance estimation algorithm, results in a comprehensive output that includes both detection results and calculated distances. At the end of the experiment, a substantial improvement in performance metrics was observed, the model achieved increases in recall (0.4%), precision (2.2%), and Mean Average Precision (mAP) (1.5%) within the 0.5 to 0.95 threshold range, while maintaining inference speeds similar to the baseline model in PASCAL VOC dataset. Besides that, distance estimation achieved an approximate average accuracy of 90% which shows the results are highly encouraging and promising. The successful integration of computer vision and distance estimation opens new possibilities for practical applications, showcasing the potential of this approach in real-world scenarios.