Energy-Efficient Deep Neural Network for Depth Prediction

Depth sensing and estimation is a key aspect of positional and navigational systems in autonomous vehicles and robots. The ability to accurately reconstruct a dense depth map of a surrounding environment from RGB imagery is necessary for successful obstacle detection and motion planning. Since deep convolutional neural networks (DNNs) have proven to be successful at achieving high accuracy rates in image classification and regression, recent work in the deep learning space has focused on designing neural networks for depth prediction applications. However, the high accuracy of DNN processing comes at the cost of high computational complexity and energy consumption, and most current DNN designs are unsuitable for low-power applications in miniaturized robots. In this project, we aim to address this gap by applying recently developed methodologies for estimating and improving the energy-efficiency of DNNs to an existing depth-prediction DNN. We envision an outcome in which the depth-prediction DNN is modified to be better suited for a specialized hardware implementation that could be integrated with a low-power visual-inertial odometry system to result in a combined navigational system for miniaturized robots.