Scholarlink

<p>This paper presents BS-CYOLOv5, a novel framework that integrates a Deep Learning (DL)-based Backtracking Search-optimized Customized YOLOv5 architecture. The key innovation of our approach lies in three fundamental contributions: (1) a deeply customized YOLOv5 architecture with Squeeze-and-Excitation attention mechanisms specifically optimized for aerial obstacle detection; (2) a novel dual-purpose Backtracking Search Algorithm that simultaneously performs hyperparameter optimization and real-time path planning; and (3) an integrated closed-loop framework that bridges the gap between perception and action in autonomous UAV navigation. 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 10,000 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, demonstrating a novel paradigm for autonomous navigation that significantly advances beyond conventional YOLOv5 modifications through its dual-optimization architecture.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhanced UAV navigation in cluttered environments via a BS-CYOLOv5 multi-sensor approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月15日

查看全文
  • Mohammed Maiza,
  • Chahira Cherif,
  • Samira Chouraqui,
  • Abdelmalik Taleb-Ahmed

摘要

This paper presents BS-CYOLOv5, a novel framework that integrates a Deep Learning (DL)-based Backtracking Search-optimized Customized YOLOv5 architecture. The key innovation of our approach lies in three fundamental contributions: (1) a deeply customized YOLOv5 architecture with Squeeze-and-Excitation attention mechanisms specifically optimized for aerial obstacle detection; (2) a novel dual-purpose Backtracking Search Algorithm that simultaneously performs hyperparameter optimization and real-time path planning; and (3) an integrated closed-loop framework that bridges the gap between perception and action in autonomous UAV navigation. 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 10,000 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, demonstrating a novel paradigm for autonomous navigation that significantly advances beyond conventional YOLOv5 modifications through its dual-optimization architecture.