Counting and locating high-density objects using convolutional neural network

Abstract

Counting and locating objects are essential in different types of applications, as they allow performance improvements in the execution of manual tasks. Deep learning methods are becoming more prominent in this type of application because they can perform good object characterizations. However, challenges such as overlapping, occlusion, scale variations and high density of objects hinder the method’s performance, making this problem remains open. Such methods usually use bounding box annotations, which hinder their performance in high-density scenes with adjacent objects. To overcome these limitations, advancing the state-of-the-art, we propose a method for counting and locating objects using confidence maps. The first application allows for the definition of a method based on convolutional neural networks that receive a Multispectral image and detect objects from peaks on the confidence map. In a second application, we insert global and local context information with the Pyramid Pooling Module, to detect different scale objects. In addition we improve the successive refinement of the confidence map with multiple sigma values in the Multi-Sigma Stage phase. In the third application of the method, we propose a band selection module to work with hyperspectral images. In the fourth application, we evaluated the proposed method on high-density objects RGB images and compared it with state-of-the-art methods: YOLO, Faster R-CNN and RetinaNet. Finally, we expanded the method by proposing a two-branched architecture enabling the exchange of information between them. This improvement allows the method to simultaneously detect plants and plantation-rows in different datasets. The results described in this thesis show that the use of convolutional neural networks and confidence maps for counting and locating objects allows high performance. The contributions of this work should support significant advances in the areas of object detection and deep learning.