Unified Object Detection, Segmentation, and Tracking Using Mask R-CNN on COCO Subset

Abstract

Object detection, instance segmentation, and multi-object tracking are important in modern intelligent vision systems applied in autonomous tasks, transportation, as well as surveillance. Nevertheless, when working in dynamic and complex conditions, classical approaches often cannot reach a compromise between accuracy of the detection, speed of the segmentation, and the real-time. Tests were done using the COCO 2017 val2017 subset that contains a number of diverse object categories with bounding box and segmentation annotations to overcome these concerns. In order to generate a specific assessment data, certain classifications such as human, automobile, bicycle, dog and truck were filtered. Scaling, normalization and data augmentation are some of the popular preprocessing techniques employed in order to enhance model resilience. The models implemented are also the Mask R-CNN which is used in detection and segmentation alongside YOLOv5, YOLOv8, and YOLOv26 to detect and track identities in real- time. Performance is measured using metrics such as mAP, iou, Precision, Recall, MOTA, MOTP and FPS. The object tracking reliability and the better accuracy-speed trade-offs in real-world conditions are the objectives of the unified framework.