Abstract:

In order to obtain better detection accuracy and faster detection speed, and ensure the safety of airport surface operation, an improved YOLOv3 algorithm was proposed in this paper, which was improved from two aspects: network structure and loss function. Firstly, the depth-wise separable convolution was used to replace the original convolution in the backbone network, and then the regression loss function of the target frame based on DIoU ratio was constructed. Taking a general airport as the research object, a surface target detection scene was built, and a training method combining migration learning and freezing training was adopted to improve the speed of surface target detection. Finally, the recognition effect of the proposed algorithm was compared with that of the traditional YOLOv3 and YOLOv4 algorithms. The results show that the detection accuracy, recall and mean average precision (mAP) of the improved YOLOv3 algorithm are 92.96%, 80.51% and 91.96%, respectively, and the graphics processing unit processing speed is 74 f/s. Compared with the traditional YOLOv3 algorithm and YOLOv4 algorithm, the performance of the improved YOLOv3 algorithm is significantly improved, which can realize the effective detection of moving targets and further ensure the operation safety of general aviation airports.

Key words: improved YOLOv3 algorithm, general aviation airports, target detection, depth-wise separable convolution, distance intersection over union(DIoU)