Automatic classification of flying bird species using computer vision techniques
Automatic classification of flying bird species using computer vision techniques
复制标题
DOI：
10.1016/j.patrec.2015.08.015
复制
发表时间：
2016-10
期刊：
影响因子：
--
通讯作者：
John Atanbori;Wenting Duan;J. Murray;Kofi Appiah;P. Dickinson
中科院分区：
文献类型：
--
作者：
John Atanbori;Wenting Duan;J. Murray;Kofi Appiah;P. Dickinson
鸟类种群被确定为重要的生物多样性指标，因此收集可靠的种群数据对生态学家和科学家来说非常重要。然而，现有的手动监测方法是劳动密集型的，耗时的，并且可能容易出错。我们工作的目的是开发一个可靠的自动化系统，能够在飞行过程中使用视频数据对单个鸟类的物种进行分类。这具有挑战性，但适合在现场使用，因为通常需要在飞行中识别，而不是在静止时。我们提出了我们的工作，它使用了一套新的和丰富的外观特征从视频分类。我们还介绍了运动特征，包括曲率和翅膀拍频。结合正态贝叶斯分类器和支持向量机分类器，我们提出了我们的外观和运动特征的实验评估在一个数据集，包括七个物种。单独使用我们的外观特征集，我们实现了92%和89%的分类率（分别使用正态贝叶斯和SVM分类器），这显着优于最近可比的最先进的系统。单独使用运动特征，我们实现了较低的分类率，但激励我们正在进行的工作，我们试图将这些外观和运动特征联合收割机，以实现更强大的分类。
Bird populations are identified as important biodiversity indicators, so collecting reliable population data is important to ecologists and scientists. However, existing manual monitoring methods are labour-intensive, time-consuming, and potentially error prone. The aim of our work is to develop a reliable automated system, capable of classifying the species of individual birds, during flight, using video data. This is challenging, but appropriate for use in the field, since there is often a requirement to identify in flight, rather than while stationary. We present our work, which uses a new and rich set of appearance features for classification from video. We also introduce motion features including curvature and wing beat frequency. Combined with Normal Bayes classifier and a Support Vector Machine classifier, we present experimental evaluations of our appearance and motion features across a data set comprising seven species. Using our appearance feature set alone we achieved a classification rate of 92% and 89% (using Normal Bayes and SVM classifiers respectively) which significantly outperforms a recent comparable state-of-the-art system. Using motion features alone we achieved a lower-classification rate, but motivate our on-going work which we seeks to combine these appearance and motion feature to achieve even more robust classification.