Echolocation using Deep Learning: An Experimental Approach

Deep Learning has found great success in the field of Computer Vision and Natural Language Processing due to the increase in computing power and data availability. Echolocation is the phenomenon of locating objects by sending out a source call in the environment and listening for echoes reflected from those objects. Mammals such as bats and whales use echolocation for navigation and prey hunting. They posses this enhanced ability where they can locate both stationary and moving objects with high precision. Echolocation can be used by autonomous systems to navigate in low light environments and can also be used in privacy constrained surveillance. In this work, we tackle a variant of echolocation where we try to predict 2D location coordinates (i.e (x,y)) of a single object. We also extend it to locate two objects. An experimental setup was designed to collect the data necessary in testing our hypothesis. A Convolution Neural Network based model is designed for object location prediction problem. The proposed model and methodology result in average prediction accuracy of (x ± 0.08, y ± 0.08) for single object detection. The current experimental design suffered from bias in samples collected and hence a few bias correcting techniques are also discussed. Finally, current challenges and a few future directions are discussed for echolocation using an experimental approach.