This thesis analyses two deep learning methods and one method with a classical approach for tracking object in video sequences, and one powerful tool for maintaining a useful appearance model. The deep learning methods are a deep regression network and a fully convolutional siamese network that have their positives and some limitations, but they still offer good performance. The method with the classic approach does tracking with kernelized correlation filters. This method uses well established theory from the field of signal processing and it derives a fast learning method for linear regression using a plethora of mathematical "tricks", like kernelization of the algorithm and clever usage of the circulant matrix properties. The method proposed for maintaining more a useful appearance model is called Tracking via Reliable Memories. It's a method that divides our appearance model into two components --- long and short term memory. It shows how fast current methods forget information, and it offers a solution to that problem. We also see that tracking participants in sporting events is a big problem for tracking systems, given the similarity between all participants (they are all human, and recordings often aren't of high enough quality) and the wealth of possible movements and deformations humans can do in those circumstances. The described methods still prove very useful for this problem, and they represent the state-of-the-art for tracking objects in video sequences.