通航机场场面运动目标检测方法

针对传统YOLOv3（you only look once-v3）算法目标检测精度较低、收敛速度较慢等问题，提出了一种改进的YOLOv3算法，分别对主干网络和损失函数进行了改进。采用迁移和冻结相结合的训练方法，以提升目标检测的精确度和速度。基于改进的YOLOv3算法对西南某通航机场3种不同场景下的运动目标检测效果进行了对比分析。结果表明，改进的YOLOv3算法对正常天气场景下的场面运动目标检测效果要明显优于雾天和雨天场景，对飞机目标的检测效果明显优于车辆和行人目标；3类目标的检测精度、召回率、平均精度值(mean average precision，mAP)分别达到92.96％、80.51％、91.96％，GPU处理速度为74f/s，较传统YOLOv3算法和YOLOv4算法性能均有明显提升。

Moving target detection method for general aviation Airport

Considering the problems of low recognition accuracy, slow training convergence of the traditional YOLOv3 algorithm, an improved YOLOv3 algorithm is proposed in this paper. The backbone network and loss function are improved respectively. In order to improve the accuracy and speed of object detection, the transfer and freeze learning strategies are adopted. Based on the improved yolov3 algorithm, the moving target detection effects of three different scenes of a general aviation airport in Southwest China are compared and analyzed. The results show that the improved yolov3 algorithm is significantly better than the fog and rain scenes in the detection of moving targets in the normal weather scene, and the detection effect of aircraft targets is significantly better than that of vehicle and pedestrian targets. The detection accuracy, recall rate and mean average precision (map) of the three types of targets reach 92.96%, 80.51% and 91.96% respectively. The GPU processing speed is 74F / s, which is significantly improved compared with the traditional yolov3 algorithm and yolov4 algorithm.