METHOD OF RECOGNITION OF MOVING OBJECTS BASED ON THE CLASSIFICATION OF HAAR CASCADES

Abstract

A method of recognition of moving objects in a video stream based on the Haar classification has been developed. When tracking objects, there is a need to identify them and record their direction of movement, speed of movement. The complexity of recognition lies not only in fixing the object and following it, but also in the movement of the camera itself, from which video surveillance is conducted. The Haar method is based on cascade classifiers that quickly highlight regions with a high probability of detecting an object. Haar cascades use a convolution operation, which is formed on the basis of the proportional product of Fourier images of functions. The disadvantages of Haar cascades include the fact that recognition is unstable when lighting changes, unstable with changes in scale and rotation of key frames. When implementing this method, no one changes the backgrounds in the video sequences. This method is very fast to implement, and accordingly the least accurate, compared to SURF and SIFT. However, it is accessible to programming and free to use. The Adaboost classifier was used to apply Haar Cascades. This algorithm selects a small number of significant features from a larger set to provide an effective result. Adaboost is an ensemble learning method that belongs to the category of boosting algorithms, which allows combining decision tree models with a small depth to create a strong model capable of providing high accuracy of classification or regression. In addition to object recognition, a machine learning method based on supervised methods was implemented to implement object location prediction and object identification. The training sample included military vehicles btr, bmp, tank, car and howitzer. It is planned to use random forest, SVM, gradient boosting and neural networks algorithms for object identification. The metrics of machine learning results are considered, in particular, the accuracy, completeness, F1-score, Kappa coefficient, and error matrix. The developed models are evaluated. In the future, it is planned to improve the methods that have been started.