Abstract:
Autonomous vehicles (AVs) have transformed the transportation sector by enhancing
safety, efficiency, and reliability through reduced human intervention. This
research presents the development and evaluation of Artificial Neural Network
(ANN)-based autonomous driving models on a scaled testbed, focusing on safety
and efficiency. The primary contribution of this research is the development of an
autonomous vehicle that achieves reliable autonomy using minimal sensor inputs
and low computational resources. The testbed relies solely on a web camera for environmental
perception and obstacle detection and uses an NVIDIA Jetson Nano
board as the computing unit. ANN-based models are developed and integrated
for lane detection, steering control, and object detection to simulate real-world
driving conditions. We evaluate two autonomous driving approaches: a modular
approach and an end-to-end learning approach. For the modular pipeline, a U-Net
model is employed for lane segmentation along with YOLOv8 (You Only Look
Once version 8) for object detection. In this approach, steering control is based
on explicit lane boundaries and detected obstacles produced by the perception
modules. In the end-to-end approach, a CNN (Convolutional Neural Network)
directly predicts the steering command from raw images, with a YOLOv8 model
running in parallel to provide obstacle awareness. Both approaches are evaluated
under different scenarios to measure their performance. To evaluate safety, testing
was conducted at three selected speeds for lane-keeping, object avoidance, and
pedestrian avoidance accuracies. Experimental results show that the U-Net-based
modular pipeline achieves a lane-keeping accuracy of 97% of lap time, while the
end-to-end CNN approach demonstrates improved obstacle and pedestrian avoidance,
achieving zero collisions in both tasks at low speeds (0.7 km/h), up to 50%
obstacle safe-zone operation, 90% pedestrian safe-zone performance, and a reduced
inference time of approximately 28 ms.
