Deep Transfer Learning-Based Obstacle Detection and Avoidance for an Autonomous Mobile Robot: A MATLAB and Simulink Implementation

Abstract

This work presents the deployment of a deep transfer learning (DTL) obstacle-detection algorithm on a four-wheel holonomic autonomous mobile robot using the Simulink and MATLAB environment. The base algorithm was constructed by fusing a convolutional neural network (CNN) with the Alex.Net transfer-learning model, and was integrated into the robot control loop through the MATLAB Robotics System Toolbox, Deep Learning Toolbox, Transfer Learning Toolbox and Neural Network Toolbox. The Simulink model interfaced a three-dimensional camera sensor block, a data-acquisition block, the DTL classifier and the kinematic/dynamic robot plant so that images acquired from the propagation path are trained, classified and mapped into steering commands in real time. The deployed system was evaluated using cross-validation on an image test-set collected on the propagation path of the robot. The DTL model achieved an obstacle detection and recognition accuracy of 98.7 %, a 1.89 % improvement over the best comparable state-of-the-art algorithm reviewed. The Simulink-in-the-loop deployment further reduced classification latency and produced a stable maneuvering response, demonstrating that the algorithm is deployable on the physical robotic platform without loss of accuracy.