Abstract Unmanned Aerial Vehicles (UAVs) are critical for applications like aerial surveillance, disaster response, and urban air mobility. Operating safely in these dynamic environments requires robust autonomous navigation and obstacle avoidance capabilities. While traditional methods depend on predefined rules and sensor-based heuristics, they often lack the real-time adaptability needed for complex scenarios. To address this limitation, we propose BS-CYOLOv5, a novel framework that integrates a Deep Learning (DL)-based Backtracking Search-optimized Customized YOLOv5 architecture. This system unifies multi-sensor data processing, real-time obstacle detection, and dynamic path planning. The model was trained and evaluated on the UAV Autonomous Navigation Dataset from Kaggle, consisting of 10000 labeled UAV flight samples gathered from RGB cameras, LiDAR, IMU, and GPS sensors. The data was split into 80% for training, 10% for validation, and 10% for testing, with min-max normalization applied to improve model performance and generalization. For obstacle detection, the CYOLOv5 model provides high-accuracy, real-time identification. The Backtracking Search Algorithm (BSA) then optimizes navigation by dynamically recalibrating flight paths for efficiency and collision avoidance while simultaneously fine-tuning the detection model's hyperparameters. Experimental results demonstrate the framework's effectiveness, showing statistically significant improvements over state-of-the-art baselines. The model achieved 98.10% obstacle detection accuracy, alongside high precision (97.52%), recall (97.54%), F1-score (97.28%), and Intersection over Union (96.1%) metrics. Comprehensive ablation studies and cross-validation confirmed the robustness of our approach. This work contributes to the advancement of intelligent UAV systems by successfully merging state-of-the-art DL with evolutionary optimization, paving the way for greater autonomy and safety in real-world operations.
Cherif et al. (Mon,) studied this question.