The proposed Confidence-Guided Cascade Decision System (CG-CDS) effectively addresses the inherent trade-off between computational efficiency and detection accuracy by dynamically routing challenging samples to a specialized fine-grained detector, which achieves robust performance across diverse environmental conditions including low illumination, severe occlusion, and complex background clutter.
Abstract
To address the challenges of construction vehicle detection caused by complex illumination and severe occlusion in transportation infrastructure scenarios, as well as the requirements for emergency rescue under sudden hazards, this paper proposes a detection system based on a multi-model decision fusion architecture. The system adopts a “coarse-screening and fine-inspection” strategy, utilizing YOLOv8 for rapid global screening to ensure real-time performance. For regions with low confidence, an improved DPFL-RetinaNet is introduced to perform secondary refined detection. Concurrently, this paper proposes the Decoupled Polynomial Focal Loss based on Taylor series expansion, which enhances the mining capability for weak targets through gradient reshaping. Experimental results demonstrate that the proposed system achieves an AP50 of 87.1% on the TCVD dataset, representing an improvement of nearly 30 percentage points over the standalone YOLOv8. This provides an efficient visual perception solution for construction safety and emergency rescue in complex environments, effectively balancing detection precision and inference speed. The proposed Confidence-Guided Cascade Decision System (CG-CDS) effectively addresses the inherent trade-off between computational efficiency and detection accuracy. By dynamically routing challenging samples to a specialized fine-grained detector, the system achieves robust performance across diverse environmental conditions including low illumination, severe occlusion, and complex background clutter. Extensive ablation studies confirm that each component of the cascaded architecture contributes positively to the overall detection performance, with the proposed Decoupled Polynomial Focal Loss (DPFL) providing the most significant improvement for weak target mining.
The proposed work introduces adaptive frame slicing scheme in the input data loader and geometric positional encoding which enables the detection of faraway vehicles with high accuracy in wide area surveillance imagery.
M. Ilamathi, Sabitha Ramakrishnan· Indian Journal of Science an...· 0 citations
With the integration of unmanned systems, aerospace remote sensing, and artificial intelligence, intelligent inspection for road infrastructure has become an important support for geohazard early warning, disaster-prevention spatial planning, and resilient urban governance. To address the slender and curved morphology of cracks, obvious scale variation, complex background interference, and missed detection of small targets in road distress detection, this paper proposes an improved YOLO11-based road distress detection method. The method designs a C3k2_DSConv module to replace the original C3k2 module of YOLO11, thereby enhancing local morphological modeling for irregular crack-like targets. A VimLayer and an AIFI module are designed and introduced to strengthen long-range continuous-structure representation and global semantic modeling. In addition, EMA attention modules are inserted before multi-scale detection features to highlight key distress regions. Experiments are conducted on the merged China and Japan subsets of RDD2022, which contain 11,753 images and 24,188 annotated bounding boxes. Experimental results show that, compared with YOLO11, the improved YOLO11 model increases mAP50 from 0.641 to 0.672 and Recall from 0.568 to 0.615. The proposed method can support UAV- and UGV-based road inspection and provide a visual perception basis for infrastructure condition assessment and resilient urban governance.
Jin-Yong Li, Jun Zhou, Can-Ming Sun et al.· 2026 3rd International Confe...· 0 citations
Drones serve as critical tools for urban security surveillance and emergency response. However, aerial object detection faces inherent challenges, as key targets such as pedestrians and vehicles often appear tiny, densely distributed, and heavily occluded. These issues reduce detection accuracy and hinder real-world security deployment. To tackle these challenges, this paper proposes WTM-YOLOv11, an improved detection model for aerial surveillance scenarios. The core improvements are threefold: 1) the WIoUv3 loss function is adopted to optimize bounding box regression and enhance the model’s adaptability to complex sample distributions; 2) we integrate a C3k2-WTConv unit (by embedding WTConv into C3k2) into the backbone to enlarge the receptive field and strengthen multi-scale feature extraction; 3) the neck network is optimized to enhance cross-scale feature fusion and small-object representation capability. Comprehensive experiments demonstrate that WTM-YOLOv11 outperforms the vanilla YOLOv11 baseline, achieving a 2.3% improvement in mean average precision (mAP) while retaining a lightweight parameter footprint. The proposed method effectively enhances the detection performance of small and occluded targets in drone aerial imagery, providing a robust and feasible solution for intelligent public security surveillance systems.
An effective real-time traffic accident detection framework based on YOLOv8 that can be implemented in intelligent transportation systems, traffic surveillance platforms, and advanced driver assistance applications is proposed.
Chuwe Ashlet Munashe, Chaoyu Yang· International Journal of Sci...· 0 citations
Traffic congestion and road safety remain major challenges for modern transportation systems, particularly in complex highway environments with high traffic volumes, tunnels, and adverse visibility conditions. Traditional monitoring methods, such as inductive loop detectors and video cameras, often suffer from high maintenance costs, limited adaptability to environmental conditions, and reduced accuracy under challenging weather or lighting conditions. To address these limitations, this paper proposes an integrated traffic monitoring framework based on omnidirectional millimeter-wave radar, high-definition road cameras, and high-precision mapping services. The proposed framework consists of a sensing layer, a data layer, and a strategy layer, enabling multi-source data acquisition, radar–camera–map data fusion, lane-level vehicle trajectory reconstruction, traffic state classification, and unsafe driving behavior identification. Field tests were conducted on the HeBa and LanHai highway sections in Guangxi Province. The results showed that the proposed system achieved high accuracy for vehicle counting and lane-change estimation, including 99.63% accuracy for car counting in the HeBa section and 98.65% accuracy for bus counting in the LanHai section. In addition, the system effectively identified unsafe driving behaviors, including stopping on the highway, wrong-way driving, slow-speed driving, and unsafe lane changes, with F1 scores ranging from 95.65% to 100% in the tested scenarios. These results demonstrate that the proposed integrated radar-based framework provides reliable lane-level traffic monitoring and safety risk identification under complex highway conditions.
Yan-Qun Yang, Xiao-Ning Deng, Shan-Feng Lu et al.· Transportation Safety and En...· 0 citations
Ship target detection technology is a key link in intelligent navigation autonomous obstacle avoidance decision-making and maritime traffic supervision, and its performance is directly related to navigation safety and the efficiency of marine governance. In the rainy conditions, due to raindrop occlusion, light scattering and other factors, the visibility is significantly reduced, which leads to the limitation of ship detection performance. Aiming at this problem, this paper improves and optimizes a object detection network based on YOLOv8s, and builds a object detection network suitable for rainy days. The improvement mainly includes two points. One is to replace the standard convolution with deformable convolution in the (Backbone) C2f module; The second is to replace bilinear interpolation up-sampling with DySample dynamic upsampling in Neck module. At the same time, to fully extend the dataset, using the DerainCycleGAN method, images of ship targets covering different rainfall intensities were synthesized and constructed. The experimental comparison between the improved model and the standard YOLOv8s model shows that the YOLOv8s-rain model has excellent detection performance, with an average accuracy rate (mAP @ 0.5) of 78.9%, which is 1.1% higher than the standard YOLOv8s model.
Hongshuai Xie, Xiu-Wen Liu, Jia-Ming Zhou et al.· International Conference on...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
