Clasificación y conteo de tráfico en video mediante un paradigma de inteligencia computacional

Abstract

In this project, an algorithm for traffic classification and counting was built using neural networks as classification model, these networks were trained using backpropagation with stochastic gradient descent and a simple genetic algorithm. The types of networks used for this were convolutional and feedforward, this last one with the idea of not leaving behind the feature engineering that might be a challenge in problems where there are not deep learning models that get the features for themselves. Once these models were obtained, the algorithm was built with the purpose of classifying a region of interest which was taken through background substraction algorithms and then adding an element which will track the vehicle through the whole window til it reaches a limit where it will be counted and added to de database having in mind characteristics as color, type of vehicle and speed estimation. The type of vehicle was classified using a neural network and the color was chosen using an euclidean distance with some base colors, finally, the speed estimation was done using the average speed formula. The final evaluation of the algorithm showed a max error of 10.6% in a test video of more than 2000 vehicles, as a final factor, a web platform was developed in order to deploy it in a device as aditional to the project as this isn't taken into account by the objectives of this, the device will be described in its own chapter.