A Precision Efficient Method for Collapsed Building Detection in Post-Earthquake UAV Images Based on the Improved NMS Algorithm and Faster R-CNN

Abstract

The results of collapsed building detection act as an important reference for damage assessment after an earthquake, which is crucial for governments in order to efficiently determine the affected area and execute emergency rescue. For this task, unmanned aerial vehicle (UAV) images are often used as the data sources due to the advantages of high flexibility regarding data acquisition time and flying requirements and high resolution. However, collapsed buildings are typically distributed in both connected and independent pieces and with arbitrary shapes, and these are generally more obvious in the UAV images with high resolution; therefore, the corresponding detection is restricted by using conventional convolutional neural networks (CNN) and the detection results are difficult to evaluate. In this work, based on faster region-based convolutional neural network (Faster R-CNN), deformable convolution was used to improve the adaptability to the arbitrarily shaped collapsed buildings. In addition, inspired by the idea of pixelwise semantic segmentation, in contrast to the intersection over union (IoU), a new method which estimates the intersected proportion of objects (IPO) is proposed to describe the degree of the intersection of bounding boxes, leading to two improvements: first, the traditional non-maximum suppression (NMS) algorithm is improved by integration with the IPO to effectively suppress the redundant bounding boxes; second, the IPO is utilized as a new indicator to determine positive and negative bounding boxes, and is introduced as a new strategy for precision and recall estimation, which can be considered a more reasonable measurement of the degree of similarity between the detected bounding boxes and ground truth bounding boxes. Experiments show that compared with other models, our work can obtain better precision and recall for detecting collapsed buildings for which an F1 score of 0.787 was achieved, and the evaluation results from the suggested IPO are qualitatively closer to the ground truth. In conclusion, the improved NMS with the IPO and Faster R-CNN in this paper is feasible and efficient for the detection of collapsed buildings in UAV images, and the suggested IPO strategy is more suitable for the corresponding detection result’s evaluation.

Keywords

• collapsed buildings detection
• UAV images
• deformable convolution
• NMS
• Faster R-CNN
• intersected proportion of objects (IPO)