A comparative analysis of recurrent deep neural networks for object tracking

Object tracking is a recurrent problem in the current world: video surveillance, autonomous driving, military scenarios, just to name a few. Several methods have been applied to this problem, such as algorithms based on hand crafted features, however deep learning methods have shown better results, due to their ability to capture complex transformations and to scale with the number of training examples. Nowadays, most of the state-of-the-art trackers use a convolutional neural network, whereas only few of them combine both a convolutional and recurrent neural network to both capture the visual information about the tracked object and remember its appearance over time. The goal of this work is to provide the reader with a comparison of the current state of the art trackers combining both convolutional and recurrent neural networks, as well as trying different architectures to further improve their performance. This work is divided in two phases: the benchmarking phase and the development phase. In the benchmarking phase we benchmarked three state-of-the-art recurrent trackers, to establish their relative performance, both in terms of accuracy and framerate, using the well-know benchmarking suite OTB, consisting of up to 100 different videos of generic object tracking. Moreover we introduced a new testing methodology which allows to test different trackers according to a detector's detections, in order to simulate a real-word scenario where the tracker is provided from the detector with a bounding box which may not perfectly enclose the tracked object as in the benchmark case. In the development phase, instead, we selected the most promising tracker of the three and performed different experiments such as fixing the weights of the convolutional neural network and trying different architectures for the recurrent neural network. The main objective of these experiment was not only to improve the model both in terms of accuracy, robustness and training time, but also to evaluate the effectiveness of different recurrent architectures and their potential use for object tracking.