BS-CYOLOv5: A Deep Learning and Backtracking Search Framework for Multi-Sensor UAV Navigation and Obstacle Avoidance

Abstract

Unmanned Aerial Vehicles (UAVs) are increasingly used in applications such as aerial surveillance, disaster response, and urban air mobility. Autonomous navigation and obstacle avoidance are critical for UAVs to operate safely in dynamic environments. Traditional obstacle avoidance methods rely on predefined rules and sensor-based heuristics, which struggle with real-time adaptability in complex environments. This presents a Deep Learning (DL)-based Backtracking Search-mutated Customized YOLOv5 (BS-CYOLOv5) framework, integrating multi-sensor data processing, real-time obstacle detection, and optimized path planning. The dataset includes 10,000 labeled UAV flight data samples recorded using RGB cameras, LiDAR, IMU, and GPS, the gathered data was split with 80% training and 20 % testing. Data pre-processing techniques, such as normalization, are applied to enhance model performance and generalization. For obstacle detection, the CYOLOv5 model is employed due to its high detection accuracy, enabling real-time identification of obstacles in various environments. Navigation and path planning are optimized using the Backtracking Search Algorithm (BSA), which dynamically adjusts flight trajectories to ensure efficient and collision-free navigation. Although these components play independent functions, they are tightly integrated into the BS-CYOLOv5 architecture, where BS also finetunes detection-related hyper parameters to improve performance. Experimental results demonstrate that the proposed approach enhances UAV obstacle avoidance accuracy and navigation efficiency. The results reveal statistically substantial improvements in the obstacle detection accuracy (98%) and high precision, recall, f1-score and IOU, when evaluated on the full 2,000-image test set. It contributes to advancing intelligent UAV systems by integrating state-of-the-art DL and optimization techniques for enhanced autonomy and safety in real-world scenarios.