For a robot to be deployed in unconstrained real world environments, it needs to be autonomous. In this preliminary work, we focus on the capacity of an autonomous robot to discover and recognize objects in its visual field. Current existing solutions mainly employ complex deep neural architectures that need to be pre-trained using large datasets in order to be effective. We propose a new model for autonomous and unsupervised object learning in videos that does not require supervised pre-training and uses relatively simple visual filtering. The main idea relies on the saliency-based detection and learning of objects considered similar (thanks to a spatio-temporal continuity). For this purpose the learning of objects is based on a Siamese Neural Network (SNN). We demonstrate the capacity of the SNN to learn a good feature representation despite the deliberately simple and noisy process used to extract candidate objects.