ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2022, Vol. 59 ›› Issue (4): 882-893.doi: 10.7544/issn1000-1239.20200986

• 人工智能 • 上一篇    下一篇



  1. 1(中国科学院计算技术研究所 北京 100190);2(中国航空无线电电子研究所 上海 200241);3(北京邮电大学信息与通信工程学院 北京 100876) (
  • 出版日期: 2022-04-01
  • 基金资助: 

A Lightweight UAV Object Detection Algorithm Based on Iterative Sparse Training

Hou Xin1, Qu Guoyuan2, Wei Dazhou2, Zhang Jiacheng3   

  1. 1(Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190);2(Chinese Aeronautical Radio Electronics Research Institute, Shanghai 200241);3(School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876)
  • Online: 2022-04-01
  • Supported by: 
    This work was supported by the National Key Basic Research and Development Program of China (2018YFC0809300, 2107YFB0202105, 2016YFB0200803, 2017YFB0202302), the National Natural Science Foundation of China (61972376), and the Beijing Natural Science Foundation (L182053).

摘要: 随着无人机技术的成熟,配备摄像头的无人机被广泛应用于各个领域,自动高效地分析和理解从无人机收集的视觉数据非常重要.基于深度卷积神经网络的目标检测算法在许多实际应用上取得了惊人的成绩,但往往伴随着巨大的资源消耗和内存占用.因此,对于无人机上携带的计算能力受限的嵌入式设备来说,直接运行深度卷积神经网络非常具有挑战性.为了应对这些挑战,以经典的目标检测方法YOLOv3(you only look once)为例,基于迭代稀疏训练的剪枝方式可以实现有效的模型压缩,同时通过组合不同数据增强方式与相关优化手段保证压缩前后检测器精度误差在可接受范围内.实验结果证明,基于迭代稀疏训练的剪枝方法在YOLOv3上取得了非常可观的压缩效果,并且将精度误差控制在了2%以内,为无人机目标检测实时应用提供了支持.

关键词: YOLOv3算法, 模型压缩, 迭代稀疏训练, 数据增强, 精度误差小

Abstract: With the maturity of UAV (unmanned aerial vehicle) technology, vehicles equipped with cameras are widely used in various fields, such as security and surveillance, aerial photography and infrastructure inspection. It is important to automatically and efficiently analyze and understand the visual data collected from vehicles. The object detection algorithm based on deep convolutional neural network has made amazing achievements in many practical applications, but it is often accompanied by great resource consumption and memory occupation. Thus, it is challenging to run deep convolutional neural networks directly on embedded devices with limited computing power carried by vehicles, which leads to high latency. In order to meet these challenges, a novel pruning algorithm based on iterative sparse training is proposed to improve the computational effectiveness of the classic object detection network YOLOv3 (you only look once). At the same time, different data enhancement methods and related optimization means are combined to ensure that the precision error of the detector before and after compression is within an acceptable range. Experimental results indicate that the pruning scheme based on iterative sparse training proposed in this paper achieves a considerable compression rate of YOLOv3 within slightly decline in precision. The original YOLOv3 model contains 61.57 MB weights and requires 139.77GFLOPS(floating-point operations). With 98.72% weights and 90.03% FLOPS reduced, our model still maintains a decent accuracy, with only 2.0% mAP(mean average precision) loss, which provides support for real-time application of UAV object detection.

Key words: YOLOv3, model compression, iterative sparse training, data enhancement, low precision loss