Deep learning based autonomous direction estimation

Outdoor mapping and localization based on appearance is especially challenging since usually separate processes of mapping and localization are required at different times of day. The problem is harder in the outdoors where continuous change in sun angle can drastically affect the appearance of a scene. In this work, we propose a method for instantaneous visual direction determination for the autonomous mobile platforms assuming the mobile platform travels along a routine route. We propose a deep convolutional neural network based algorithm for classification of instantaneous images of the path to be followed. The model is tested on SeqSlam dataset and a success performance of %78.5 is achieved. Hidden layer weights are analyzed to ensure that the learning is actually achieved. Experimental results suggest that deep neural networks yield high recognition rates of images to be used for autonomous movement. Approach will be tested on a novel dataset and its performance will be realized in realtime as future work.