SEGMENTATION RECOGNITION OF CLUSTERS OF MECHANIZED OBJECTS IN AN IMAGE USING TWO-STAGE SEGMENTATION METHODS BASED ON R-CNN

Abstract

The paper addresses the critical task of segmentation recognition of clusters of mechanized objects in images using two-stage segmentation methods based on R-CNN architectures. Clusters of vehicles or other mechanized objects pose significant challenges in various fields, from traffic congestion management and accident analysis to targeted operations involving UAVs. Precise identification of object boundaries and estimation of the number of objects are essential for effective decision-making in such scenarios. The authors propose a solution based on the application of Mask R-CNN and Cascade Mask R-CNN. The research includes an extensive review of recent studies that applied Mask R-CNN and ResNet-50 architectures for aerial and object detection tasks, highlighting their strengths and limitations, particularly under varying visibility conditions. 
In the experimental part, the results demonstrate that the cascade R-CNN mask outperforms the R-CNN mask in general, achieving higher precision across most categories. However, in specific cases at high IoU thresholds, Mask R-CNN showed competitive or superior performance. Despite Cascade Mask R-CNN’s overall advantages, including improved segmentation accuracy due to its cascade structure refining detection at multiple IoU thresholds, computational complexity and hardware requirements remain significant challenges. 
Future work may focus on optimizing model inference speed and robustness under adverse environmental conditions to facilitate broader real-world deployment.