Multi-Channel Gaussian Derivative Neural Networks for Crowd Analysis
Hugo Gavilima-Pilataxi, Julio Ibarra-Fiallo · 2023
This research shows the procedure to replace the image filtering for the counting of individuals carried out with a Gaussian filter kernel in order to obtain a density value (number of individuals) in a crowd, with Multi-Channel Gaussian Derivative Neural Networks. Gaussian operators, based in Scale-Space Theory, allows processing visual information in greater detail, especially in data sets for crowd counting with different scales, occlusion problems, or complex scenarios, which results in perfect candidates to be used as a primitive structure in a layer in deep neural network to significantly reduce the number of parameters in the model. Overall, the proposed mode achieves metrics comparable to high-level models, while using only approximately 10% of the parameters, which suggests a possible solution or future line of research for the study of urban congestion. In this way, Gaussian derivative neural network allows for more efficient processing of visual information and reduces the number of parameters required, making it an attractive option for crowd analysis in urban areas.