Approche de suivi visuel d'objet basée sur la Transformée de Cosinus Discret DCT et les reseaux de neurones convolutionnels CNN

Abstract

This thesis presents a groundbreaking Visual Object Tracking (VOT) approach designed to tackle a prevalent challenge in existing methods: the considerable alteration in object appearance, primarily stemming from extensive occlusion and varying illumination conditions. The proposed method integrates several key components, including Deep Convolutional Neural Networks (DCNN), Discrete Cosine Transform (DCT), Histograms of Oriented Gradients (HOG) features, and an HSV-based energy condition. Initially, an HSV-based energy condition enriches the learning process by merging both RGB and HSV color bases, enhancing the model's adaptability. Rather than relying on the image template, this technique utilizes the coefficients derived from the image's DCT to handle high saturation images in the Convolutional Neural Networks (CNN's) input. Extracting CNN features involves utilizing the Inverse Discrete Cosine Transform (IDCT). Subsequently, the approach harnesses multichannel correlation maps generated by CNNs to precisely determine the target's position. This is achieved through the amalgamation of convolutional features. Newton's method plays a pivotal role in this process, bolstering the system's long-term memory regarding the target's appearance and aiding in recovery from tracking failures. Further, the updating parameter for the correlation filters is determined by selecting the highest value among the output maps derived from correlation filters using convolutional features extracted from the HOG features of the image template. The conclusive results unequivocally establish the superiority of the proposed method, surpassing the performance of most recent tracking techniques.