Spatial instance segmentation in monocular images through convolutional neural networks

Abstract

The driver assistance systems for partial or full automation of the vehicle are systems that help in decision making and have an important role in traffic safety and efficiency. Such systems require robust perception system that should be able to handle occlusions and complex urban scenarios in order to allow the vehicle understanding about the environment where it is navigating. Therefore, this doctoral project proposes a convolutional neural network based method for spatial instance segmentation, such that the instance segmentation and depth information are estimated simultaneously from a monocular image. The extraction of features relevant to tasks learning will be performed by a convolutional neural network, which will be designed as an encoder-decoder architecture for pixel-level classification and multi-task learning. This architecture will consist of a unique encoder and split decoders. Aiming better performance of the network, two research objects will be studied, the set of loss functions that provides the best convergence during training and the architecture design that favors parameters sharing between tasks' decoders during the learning process. For validation, the performance of the proposed method will be compared in existing benchmarks. Furthermore, real experiments will be executed within the CaRINA II platform, an ongoing autonomous vehicle project held by the Mobile Robotics Laboratory (LRM) with Intelligent Systems Laboratory (LASI) of the University of São Paulo at São Carlos. This project is part of the thematic project supported by FAPESP, process n. 2014/50851-0 "National Institute of Science and Technology for cooperative autonomous systems applied to security and environment". (AU)