Real-Time Obstacle Detection Algorithm based on Visual Perception in Unmanned Vehicles

Abstract

This study addresses the issue of identifying obstacles in real time for unmanned automobiles using visual perception, which is a crucial component for ensuring the safety of autonomous driving. Existing detection methods are often limited in accuracy and robustness under diverse lighting and weather conditions. The purpose of this project is to create a new model that enhances detection precision, adaptability and immediate efficiency over a wide range of traffic circumstances. By making use of the KITTI dataset, which is divided into the training, validation and test sets in a ratio of 7:2:1, the performance of the proposed model is evaluated on the basis of detection efficiency, recall rate and average precision. These results are then compared to the results of YOLOv4, Faster Region-based Convolutional Neural Network (R-CNN) and classic convolutional neural network models. The results of the experiments show that, in bright settings, the model attains a median detection accuracy of 0.9275. In addition, it obtains a pedestrian detection accuracy of 0.88 and a vehicle detection accuracy of 0.90 in rainy situations. Furthermore, the model has achieved remarkable accuracy in detecting obstacles at night. Additionally, it demonstrates improved memory rates and average precision in circumstances where traffic is congested. The exceptional performance of the model can be ascribed to its distinctive modular configuration and mechanisms for interaction that are highly efficient. This research offers a viable and effective approach for real-time obstacle detection in autonomous vehicles, despite the fact that restrictions exist owing to experimental settings and the dataset.