Accurate displacement information is of great significance for the structural safety state assessment. Due to the advantages of remote, high-resolution and easy implementation, the vision-based monitoring technologies...
详细信息
Accurate displacement information is of great significance for the structural safety state assessment. Due to the advantages of remote, high-resolution and easy implementation, the vision-based monitoring technologies have received widespread attention in the field of structure health monitoring. However, these are highly sensitive to lighting changes and complex scenarios in monitoring environment, which affects their further development in practical applications. Traditional monitoring methods typically rely solely on low-level appearance features to identify monitoring objects under complex scenarios. However, these features are not sufficient to achieve satisfactory accuracy, especially for low resolution images. The deep learning network based on YOLO series exhibits robust adaptability in challenging scenarios. This paper proposed an improved YOLOv8 structure for object detection under complex scenarios, where the global attention module (GAM),adaptive feature fusion (AFF),and diverse branch cross stage partial (DBCSP) modules are designed. AFF module effectively preserves the complex features of small objects in complex scenarios. DBCSP module introduces a multi-branch strategy to enhance the learning and presentation ability of object feature. Then, the state-of-the-art bytetrack algorithm is introduced to track the detection objects for the aim of displacement extraction. Finally, the self-made dataset with circular codes and the cantilever beam with circular coded objects on its surface were used to validate the proposed structural displacement monitoring method. The dataset takes into account different complex backgrounds, scales or shooting distances, shooting angles and lighting intensities, and data augmentation techniques are used to improve the generalizability of model. The experiment shows that the improved YOLOv8 structure improves the average precision (AP) by 5.8 % on the dataset compared with YOLOv8 structure. Although the improved YOLOv8 and Y
This paper focuses on the improvement of kernel filtering algorithms and their applications, particularly on enhancing the accuracy of target recognition and tracking in modern intelligent environmental monitoring. Th...
详细信息
ISBN:
(纸本)9798350389814;9798350389807
This paper focuses on the improvement of kernel filtering algorithms and their applications, particularly on enhancing the accuracy of target recognition and tracking in modern intelligent environmental monitoring. The Kernel Correlation Filter (KCF) algorithm, due to its efficiency, high precision, and strong robustness, holds significant potential for application in surveillance environments. However, the performance of the KCF algorithm is limited in complex scenarios such as occlusion, scale changes, similar targets, and rapid movements. To address these issues, this paper proposes two improved KCF algorithms. Firstly, a scale pool algorithm is introduced, which enhances the algorithm's adaptability to target scale changes by performing multi-scale feature extraction and matching, thereby improving tracking accuracy, efficiency, and speed. Experiments demonstrate that the tracking accuracy of the improved algorithm is significantly enhanced on the OTB50 and OTB100 datasets. In experiments using a self-constructed dataset, the success rates for feature extraction, scale changes, and sampling rates were improved by 40%, 13.6%, and 29.76%, respectively. Secondly, to address the occlusion problem, an improved joint tracking algorithm combining the modified KCF with bytetrack is proposed. The high-quality detection results of the improved KCF are utilized as input for bytetrack, which optimizes the handling of occluded objects through its multiple matching strategies. The joint algorithm exhibits excellent performance on the OTB50 and OTB100 datasets. In the self-constructed dataset, the success rate for handling occlusion scenarios is improved by 50% compared to the original KCF, effectively solving issues such as target template updating, feature changes, and target disappearance.
暂无评论