Glimpse-gaze deep vision for Modular Rapidly Deployable Decision Support Agent in smart jungle

Visual interpretation of complex visual patterns in non-urban environments is necessary for many applications in smart rural community management, smart farming and smart jungles. In this paper, the Glimpse-Gaze framework for deep learning based visual interpretation of complex rural and jungle environment scenes is proposed. The proposed framework is used for decision support and navigation by a multi-agent robotic system singularly referred to as MOdular RApidly Deployable Decision Support Agent (MORAD DSA). A set of deep con-volutional neural networks are trained for fast and accurate interpretation of jungle scenes. Transfer learning and auxiliary pretraining on salient regions of the jungle scenes are investigated and the hyper parameter tuning and data augmentation for avoiding overfitting for the proposed model are explored. The experimental results show that the Glimpse-Gaze framework is capable of generating accurate visual cues for precise navigation and visual interpretation in the unstructured rural and jungle environments. A series of data sets for smart jungle applications are collected and the proposed framework is evaluated for applications such as detection of fire hazards and illegal grazing in the jungle.