光学学报, 2020, 40 (19): 1915001, 网络出版: 2020-10-12
面向无人机的轻量级Siamese注意力网络目标跟踪 下载: 1599次
Light-Weight Siamese Attention Network Object Tracking for Unmanned Aerial Vehicle
机器视觉 目标跟踪 Siamese网络 MobileNet 通道注意力 空间注意力 协同注意力 machine vision object tracking Siamese networks MobileNet channel attention spatial attention coordination attention
摘要
随着无人机技术在**、民用等领域的广泛运用,高精度、低功耗智能无人机跟踪系统的需求也日益增多。针对无人机跟踪任务中目标尺度变化大、视野角度多变、遮挡等问题,提出了一种基于轻量级Siamese注意力网络的无人机实时跟踪算法。首先,选取易于部署在嵌入式设备中的轻量级卷积神经网络MobileNetV2作为特征提取主干网络;接着,设计通道空间协同注意力模块,增强模型的适应能力与判别能力;然后,搭载区域建议网络,通过互相关获取前景背景分类和边界框回归响应图;最后,加权融合多层响应图,调整候选区域筛选策略,计算得到更加准确的跟踪结果。在无人机跟踪数据集上的仿真实验结果表明,相对于当前主流算法SiamRPN,该算法跟踪精度提升了3.5%,能更好地应对复杂多变的场景。同时,在NIVIDA RTX 2060 GPU上,跟踪速度达到60 frame/s。
Abstract
With the widespread use of unmanned aerial vehicle (UAV) technology in military, civilian, and other fields, the demand for high-precision, low-power intelligent UAV tracking systems is also increasing. Aiming at the problems of scale variation, out-of-view, and occlusion in UAV tracking tasks, a real-time tracking algorithm for UAV based on light-weight Siamese network was proposed. Firstly, the lightweight convolutional neural network MobileNetV2, which is easy to be deployed in embedded devices, is selected as the feature extraction backbone network. Secondly, the channel spatial coordination attention module is designed to enhance the adaptive and discriminative ability of the model. Thirdly, the region proposal network is equipped, and the foreground background classification and boundary box regression response map are obtained through correlation. Finally, the weighted fusion multilayer response map is calculated and proposal region screening strategy is adjusted to obtain more accurate tracking results. Simulation experimental results on the UAV tracking dataset show that the tracking accuracy is improved by 3.5% compared to the current mainstream algorithm SiamRPN, and the algorithm can better cope with complex and changeable scenes. Meanwhile, on the NIVIDA RTX 2060 GPU, the tracking speed achieves 60 frame/s.
崔洲涓, 安军社, 张羽丰, 崔天舒. 面向无人机的轻量级Siamese注意力网络目标跟踪[J]. 光学学报, 2020, 40(19): 1915001. Zhoujuan Cui, Junshe An, Yufeng Zhang, Tianshu Cui. Light-Weight Siamese Attention Network Object Tracking for Unmanned Aerial Vehicle[J]. Acta Optica Sinica, 2020, 40(19): 1915001.