Research On Crowd Counting Algorithm Based On Deep Learning

As a common task in people’s daily life,object counting is a challenging problem in computer vision.It is difficult to achieve accurate result on crowd counting due to inter-object occlusion problems,multi-scale problems and background diversity.Most of the existing work on crowd counting revolves around crowds,and most of the work related to crowd counting is performed based on public datasets.In order to improve the accuracy of existing crowd counting algorithms,this thesis focused on how to improve the accuracy of sturgeon counting in real farming scenarios and how to improve the counting performance of existing crowd counting algorithms on public datasets.The main work and contributions of this thesis can be summarized as follows:1.Construction of sturgeon dataset.In order to facilitate the subsequent sturgeon counting research,a sturgeon dataset was constructed in this thesis under real farming scenarios.The whole process of dataset construction includes data collection,preprocessing and labeling work as well as density map generation.The constructed dataset played a key role in the subsequent evaluation of the sturgeon counting algorithm.2.Sturgeon counting based on Image processing.In order to reduce the cost of sturgeon counting,this thesis achieves accurate sturgeon counting by image processing methods such as histogram equalization,circle detection,grayscale map binarization and contour detection.Compared with deep-learning based methods,this method does not require additional model training and data annotation but can achieves comparable sturgeon counting accuracy.3.Sturgeon counting based on deep learning.To improve the robustness of the sturgeon counting algorithm,this thesis introduces deep learning into the sturgeon counting process.After realizing accurate sturgeon counting based on existing crowd counting algorithms,this thesis proposed a sturgeon counting algorithm based on weakly supervised learning and Transformer,which improved the sturgeon counting accuracy and reducing the cost of data labeling.4.Crowd counting based on collaborated attention mechanism.To further improve the counting accuracy of existing crowd counting algorithms on public datasets,this thesis proposed a dual channel attention module and a multi-branch window learnable region attention module based on the baseline respectively,which effectively reduced the counting error of the baseline and achieved better crowd counting performance by combining the channel attention mechanism with the multi-scale spatial attention mechanism.