Comparative Analysis of Machine Learning Models for Vision-Based Line Following Rovers

In recent years, machine learning (ML) has advanced autonomous systems, allowing for more dynamic methods of implementing line following robots using convolutional neural networks (CNNs) to process visual data. This paper presents a line-following rover for autonomous navigation, leveraging machine learning to interpret visual input. The proposed design investigates several ML models for line classification, using a camera for visual input and a Proportional-Integral-Derivative (PID) controller for motor control. The design uses a Raspberry Pi 4B and the Sphero RVR. Five different models—DenseNet121, EfficientNet-B0, EfficientNet-B1, MobileNet, and SqueezeNet—were assessed and compared for their effectiveness in line classification and cycle computation time. The dataset was constructed by manually guiding the rover along a track generating a total of 4,378 images. Experimental findings showed that lightweight models like SqueezeNet performed better compared to larger models such as DenseNet121, with cycle computation times of 201.87ms and 271.73ms, respectively.