The appearance-based approach in visual robot navigation supposes several advantages, such as its application to non-structured environments and the relatively simple extraction of control laws that it offers. However, the main drawback is the requirement of extensive memories and the high computational cost. This way, the nature and the quantity of information to store about the environment is very important. This work presents how to reduce the dimension of the database, by means of calculating just the most significant information of each image. We show how it can be done working in the PCA subspace. This method allows lowering the computational cost without necessity of reducing the resolution of the images, what implies that it could be used in very non-structured environments, in the presence of partial occlusions and with considerably high translational speed of the robot.