DGR Days 2016

Deep Convolutional Neural Network (DCNN) models have achieved unprecedented performance on various image understanding tasks. However, these models are still susceptible to several disturbances that affect perception in real-world environments including lighting changes, glare and ambient noise. In addition, training of DCNNs require a large amount of annotated data which is difficult to acquire for specialized robotic applications. In this paper, we describe two of our systems that overcome some of these limitations by learning end-to-end hierarchical noise-robust features. First, we give an overview of our multispectral DCNN fusion architecture that enables the network to learn features from multiple modalities and spectra for dense semantic segmentation of forested environments. Secondly, we describe our proprioceptive spatiotemporal model for terrain classification from acoustic vehicle terrain interaction signals. Both of our approaches achieve state-of-the-art performance on extensive realworld data compared to existing techniques. Moreover, our networks perform inference in real-time enabling quick decision making, critical for operating in such environments.