基于嵌入式平台的航拍目标智能识别
2022,30(11):153-160
摘要：基于多旋翼无人机实现目标识别具有成本低、灵活性高的优点,能够对近地低空目标进行高强度监测,在国防军事领域和民用领域具有巨大的应用前景；但无人机机载计算机常使用功耗小、重量轻、可靠性高的嵌入式设备,该类设备算力有限,难以实时运行现有深度学习目标识别算法,因此研究深度学习航拍小目标识别技术在嵌入式设备中实时运行有重要意义；基于YOLOv4设计了适用于无人机俯视小目标的轻量化网络,并基于BN层 系数对网络进行剪枝,采用了TensorRT对算法进行硬件加速；同时,制作了小型军用目标数据集,基于该数据集,在机载嵌入式运算平台上对原始YOLOv4算法和改进的算法分别进行了测试,改进算法与原YOLOv4相比,准确率提升了2.3%,速度提升了3.3倍。
关键词：目标识别;YOLOv4;深度学习;无人机;网络轻量化;
The Target Intelligent Recognition of Aerial Photography Images Based on Embedded Platform
Abstract：Target recognition and tracking based on multi-rotor UAVs has the advantages of low cost and high flexibility. It can carry out high-intensity monitoring of low-ground and low-altitude targets. It has a huge application prospect in the national defense and military fields and civilian fields. Embedded devices with low power consumption, light weight and high reliability are often used in UAV airborne computers. The computing power of such devices is limited, and the existing target recognition algorithms based on deep learning are difficult to run on such devices in real time. Therefore, it is of great significance to research on the technology of aerial small target recognition based on deep learning and running real time in embedded equipment. A lightweight YOLOv4 backbone network which suitable for overlooking small targets is improved and designed, and the network is pruned based on the γ coefficient of the BN layer, and accelerated using TensorRT. Besides, an image dataset of military target is produced, and it is used for testing the improved algorithm on the airborne embedded computing platform. Compared with the original YOLOv4, the accuracy of the improved algorithm is increased by 2.3% and the speed is increased by 3.3 times.
Key words：Target recognition; YOLOv4; deep learning; UAV; network lightweight
收稿日期：2022-08-05
基金项目：国家自然科学基金(61973160,62073161),江苏省自然科学基金(BK20210298)
