无人机对地小目标检测方法研究

Research on detection methods of small targets on ground by UAV

无人机图像中的小目标检测是研究的重难点之一。与大目标相比,小目标特征较少,更易受到遮挡和复杂背景的干扰,针对该问题,提出一种基于YOLOv7-tiny的多模型融合目标检测网络YOLO-DA。首先,增加小目标和极小目标检测层,提高网络对小目标特征的学习能力;其次,引入空间自适应特征融合ASFF-L检测头,通过学习空间过滤冲突信息来抑制不同尺度特征的不一致性,实现多尺度特征的自适应融合;最后,引入DCNS可变形卷积并设计了扩展变形建模范围的调制机制,增强模型的建模能力,降低遮挡重叠等对检测的影响。经试验验证,提出的方法在Visdrone2019数据集上实现了44.7%的平均精度及71 fps的推理速度,平均精度较基线算法提高了9.7%,模型内存为63.8 M,能够实现实时检测。通过消融、对比实验表明YOLO-DA在无人机航拍图像检测方面明显减少了误检和漏检问题,具有更高的检测性能,且算法参数量和计算量可以满足无人机等边缘设备的实时检测需求。

Small object detection in drone images is one of the key and difficult research areas.Compared with large targets,small targets have fewer features and are more susceptible to interference from occlusion and complex backgrounds.To address this issue,a multi model fusion object detection network YOLO-DA based on YOLOv7 tiny is proposed.Firstly,add layers for detecting small and extremely small targets to enhance the network’s ability to learn small target features;Secondly,the spatial adaptive feature fusion ASFF-L detection head is introduced to suppress the inconsistency of features at different scales by learning spatial filtering conflict information,achieving adaptive fusion of multi-scale features;Finally,DCNS deformable convolution was introduced and a modulation mechanism was designed to expand the range of deformable modeling,enhance the modeling ability of the model,and reduce the impact of occlusion overlap on detection.Through experimental verification,the proposed method achieved an average accuracy of 44.7%and a inference speed of 71 fps on the Visdrone2019 dataset.The average accuracy was improved by 9.7%compared to the baseline algorithm,and the model memory was 63.8 M,enabling real-time detection.Through ablation and comparative experiments,it has been shown that YOLO-DA significantly reduces false positives and false negatives in drone aerial image detection,and has higher detection performance.Moreover,the algorithm parameters and computational complexity can meet the real-time detection requirements of edge devices such as drones.