MEConvNN-Designing Memory Efficient Convolution Neural Network for Visual Recognition of Aerial Emergency Situations

Abstract

Unmanned aerial vehicles (UAVs) play a vital role in calamity and natural disaster management due to their remote sensing capabilities. Specifically, UAVs/drones equipped with visual sensors can have remote access to confined areas wherein human access is limited. Growing inventions in deep learning incubate the efficacy of such UAVs/drones in terms of computational capability with limited resources and lead us to effectively utilize this technology in visual recognition of emergency situations, like floods in urban areas, earthquakes or fires in forests, and traffic accidents on busy highways. This can be beneficial in mitigating the consequences of such events on the environment and people more rapidly with minimum men and material loss. However, most deep learning architectures used in this domain with high accuracy are costly regarding memory and computational resources. This motivates us to propose a framework that can be computationally efficient and can be utilized on an embedded system suitable for smaller platforms. In this work, we formalize and investigate that problem and design a memory-efficient neural network for visual recognition of emergency situations named MEConvNN. To this end, we have effectively used dilated convolutions to extract the spatial representation. The proposed method is experimentally evaluated using Aerial Image Database for Emergency Response (AIDER), showing comparative efficacy with the state-of-the-art methods. Specifically, the proposed method achieves accuracy with less than a 2% drop compared to state-of-art methods but is more memory efficient in contrast to state-of-art methods.