| dc.description.abstract | Estimating density maps and counting the number of objects of interest from images has a wide range of applications, such as crowd counting, traffic monitoring, cell microscopy in biomedical imaging, plant counting in agronomy, as well as environmental survey. Manual counting is a labor-intensive and time-consuming process. Over the past few years, the topic of automatic object counting by computers has been actively evolving from the classic machine learning methods based on handcrafted image features to end-to-end deep learning methods using data-driven feature engineering, for example by Convolutional Neural Networks (CNNs). In our research, we focus on the task of counting plants for large-scale nursery farms to build an AI-horticulture monitoring and prediction system using unmanned aerial vehicle (UAV) images. The common challenges of automatic object counting as other computer vision tasks are scenario difference, object occlusion, scale variation of views, non-uniform distribution, and perspective difference. For an AI-horticulture monitoring and prediction system for large-scale analysis, the plant species various a lot, so that the image features are different based on different appearance of species. In order to solve these complex problems, the deep convolutional neural network-based approaches are proposed.
Our method uses the density map as the ground truth to train the modified classic deep neural networks for object counting regression. Experiments are conducted comparing our proposed models with the state-of-the-art object counting and density estimation approaches. The results demonstrate that our proposed counting model outperforms state-of-the-art approaches by achieving the best counting performance with a mean absolute error of 1.93 and a mean square error of 2.68 on our horticulture nursery plant dataset. | en |
