Remote Sensing-based Vehicle Monitoring System using YOLOv10 and CrossViT

Rapid urbanization and vehicular traffic growth have significantly strained traditional traffic management systems, exacerbating issues such as congestion, safety hazards, and environmental degradation. This paper presents a novel deep learning-based framework designed to address these challenges by enhancing vehicle detection, tracking, and classification in urban traffic settings. The proposed system utilizes Gaussian Blur for effective image preprocessing, followed by DeepLabv3 for accurate vehicle segmentation. Vehicle detection is performed using the YOLOv10 architecture, which excels in detecting vehicles in complex and occluded environments. For vehicle counting, we employ CenterNet, an efficient method that leverages keypoint detection to estimate vehicle counts. Optical flow techniques are applied to track vehicle motion and improve counting accuracy in dynamic traffic conditions. The system’s feature extraction pipeline incorporates deep convolutional networks to capture critical vehicle characteristics, while vehicle classification is enhanced by the CrossViT model, which excels at handling varying vehicle types and scales. Evaluation on the VAID dataset reveals a vehicle detection accuracy of 90%, tracking accuracy of 88%, and classification accuracy of 91%. These results underscore the potential of the proposed framework to offer a scalable, efficient solution for intelligent transportation systems, enabling comprehensive urban traffic monitoring and management.