《计算机应用》唯一官方网站

• •    下一篇

基于嵌入式Jetson TX2的高原鼠兔目标检测

陈海燕,贾明明,赵文力,王婵飞   

  1. 兰州理工大学 计算机与通信学院,兰州 730050
  • 收稿日期:2021-11-02 修回日期:2021-12-20 接受日期:2021-12-23 发布日期:2021-12-31 出版日期:2021-12-31
  • 通讯作者: 陈海燕
  • 基金资助:
    国家自然基金项目

Target detection of Ochotona curzoniae based on embedded Jetson TX2 

  • Received:2021-11-02 Revised:2021-12-20 Accepted:2021-12-23 Online:2021-12-31 Published:2021-12-31

摘要: 高原鼠兔目标检测是对其进行种群数量统计及种群动态变化研究的基础,而传统高原鼠兔智能监测系统中往往目标检测硬件设备大,在抽样采集数据时移动性较弱。针对此问题,本文提出一种便携式的高原鼠兔目标检测方法。首先,以MobileNet替换YOLOv3主干网络DarkNet53来构建轻量级高原鼠兔目标检测模型;其次,利用剪枝、微调的方法对模型进一步轻量化;最后,将剪枝、微调后的模型部署到便携式目标检测设备Jetson TX2上。对自然场景下高原鼠兔目标检测的实验结果表明:本文方法的平均检测精度、检测速度和模型大小分别为97.36%、36帧/秒和14.88M(MByte),优于未裁剪的模型(平均检测精度:97.30%,检测速度:14帧/秒,模型大小:92.8M);同时,与原始(检测精度:98.41%,检测速度:5帧/秒,模型大小:235M)的目标检测模型相比,本文方法在平均精度下降1.05%的情况下,检测速度提升了620%,模型大小压缩了93.7%;与现有轻量级模型相比,在平均检测精度、检测速度、模型大小指标中,均有明显提升。

关键词: 高原鼠兔, Jetson TX2, 轻量化, 目标检测, 模型剪枝

Abstract: Target detection of Ochotona curzoniae is the basis of population statistics and population dynamics. In the traditional intelligent monitoring system of Ochotona curzoniae, the target detection hardware equipment is often large and has weak mobility when sampling and collecting data. To solve above problems, a portable target detection method of Ochotona curzoniae was proposed. Firstly, a lightweight Ochotona curzoniae target detection model was constructed by replacing YOLOv3 backbone network Darknet53 with MobileNet. Secondly, the model was further lightweight by pruning and fine-tuning. Finally, the pruned and fine tuned model was deployed on the portable target detection device Jetson TX2. The experimental results of Ochotona curzoniae target detection in natural scene show that the average detection accuracy, detection speed and model size of the proosed method are 97.36%, 36 frames per second and 14.88M(MByte) respectively, which is better than the uncut model (average detection accuracy: 97.30%, detection speed: 14 frames per second, model size: 92.8M). At the same time, compared with the original (detection accuracy: 98.41%, detection speed: 5 frames per second, model size: 235M), when the average accuracy of the proposed method is reduced by 1.05%, the detection speed is improved by 620% and the model size is compressed by 93.7%. Compared with the existing lightweight models, the average detection accuracy, detection speed and model size are significantly improved. 

Key words: Ochotona curzoniae, Jetson TX2, lightweight, target detection, model pruning

中图分类号: