Computers, materials & continua/Computers, materials & continua (Print) · 2025 · 15 citations · 16 references
ComputingComputer ScienceComputer VisionMachine LearningSensor FusionDeep LearningImage ClassificationObject DetectionPattern RecognitionRemote Sensing Image AnalysisScene UnderstandingConditional Random Field SegmentationConditional Random FieldsContour DetectionFeature Pyramid NetworksMulti-scale Object DetectionTexture EncodingVehicle DetectionViewpoint Invariance
Unmanned Aerial Vehicles (UAVs) are increasingly employed in traffic surveillance, urban planning, and infrastructure monitoring due to their cost-effectiveness, flexibility, and high-resolution imaging. However, vehicle detection and classification in aerial imagery remain challenging due to scale variations from fluctuating UAV altitudes, frequent occlusions in dense traffic, and environmental noise, such as shadows and lighting inconsistencies. Traditional methods, including sliding-window searches and shallow learning techniques, struggle with computational inefficiency and robustness under dynamic conditions. To address these limitations, this study proposes a six-stage hierarchical framework integrating radiometric calibration, deep learning, and classical feature engineering. The workflow begins with radiometric calibration to normalize pixel intensities and mitigate sensor noise, followed by Conditional Random Field (CRF) segmentation to isolate vehicles. YOLOv9, equipped with a bi-directional feature pyramid network (BiFPN), ensures precise multi-scale object detection. Hybrid feature extraction employs Maximally Stable Extremal Regions (MSER) for stable contour detection, Binary Robust Independent Elementary Features (BRIEF) for texture encoding, and Affine-SIFT (ASIFT) for viewpoint invariance. Quadratic Discriminant Analysis (QDA) enhances feature discrimination, while a Probabilistic Neural Network (PNN) performs Bayesian probability-based classification. Tested on the Roundabout Aerial Imagery (15,474 images, 985K instances) and AU-AIR (32,823 instances, 7 classes) datasets, the model achieves state-of-the-art accuracy of 95.54% and 94.14%, respectively. Its superior performance in detecting small-scale vehicles and resolving occlusions highlights its potential for intelligent traffic systems. Future work will extend testing to nighttime and adverse weather conditions while optimizing real-time UAV inference.
16
Focal Loss for Dense Object Detection
Tsung-Yi Lin, Priya Goyal, Ross Girshick et al. · IEEE Transactions on Pattern Analysis and Machine Intelligence · 2018 · 9.6K citations
EfficientDet: Scalable and Efficient Object Detection
Mingxing Tan, Ruoming Pang, Quoc V. Le · 2020 · 8.4K citations
Zhipeng Deng, Hao Sun, Shilin Zhou et al. · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2017 · 197 citations