Enhancing Object Mapping in SLAM using CNN

Abstract

Automation is becoming more prevalent among manufacturing and eCommerce companies as a way  to better serve their customers. One of the key problems in warehouse management is controlling the internal delivery/movement of goods/objects. It is labor-intensive, time-consuming, and needs additional care based on delicacy goods. Automated guided vehicles (AGVs) that are small in size can serve as a solution to the aforementioned problem of locomotion. For any robot to move autonomously, the initial and critical requirement is to understand the surrounding environment precisely. Simultaneous Localisation and Mapping (SLAM) is the preferred method to build an environment map at runtime. SLAM is designed to work in a static environment and faces a few challenges once it involves dynamic objects. This research proposes Deep Learning to enhance the SLAM technique. It aids the identification of static and dynamic objects and consequently updates the occupancy grid map. The proposed approach has been validated through a simulated environment and a Convolution Neural Network (CNN) for the classification of static and dynamic objects. The simulation results demonstrate the promising nature of the proposed approach.