Mar 2026· 2026 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)· pp. 603-612· 1 citation· 29 references
Computer Science
Abstract
Illegal waste dumping poses significant environmental and public health challenges worldwide, requiring automated surveillance systems for detection and prevention. This paper presents our solution for the IWDD 2026 Contest, addressing the dual challenge of detecting illegal dumping events in surveillance videos and localizing the exact moment of occurrence. We employ X3D-M, an efficient 3D convolutional network pretrained on Kinetics-400, combined with a sliding window inference strategy for temporal localization. Through systematic hyperparameter optimization across 96 configurations and ablation studies examining nine combinations of fine-tuning strategies and loss functions, we identify key design choices for this application domain. Our experiments reveal that differential learning rates-applying lower rates to the pretrained backbone while training the classifier more aggressively-outperform both frozen backbones and uniform fine-tuning. The optimal system achieves an F1-score of 0.8387 and a temporal F1score of 0.7742 on our test set, with 92.3% of correct detections within the temporal tolerance window. Operating at over 8 times real-time speed with only 2.97M parameters, our approach demonstrates that efficient video classification architectures can be effectively adapted for specialized surveillance applications through careful transfer learning and inference design.
Detecting oil spills at sea is a difficult vision problem. Slicks typically have soft edges, scatter into irregular patches, and resemble several harmless features of the sea surface. This paper presents a controlled benchmark of six YOLO-based object detectors applied to this task, namely YOLOv8n, YOLOv8s, YOLO11n, YOLO11s, YOLO26n and YOLO26s. Every model was trained and evaluated under strictly identical conditions, using an annotated oil spill dataset and the AdamW optimizer. To provide a deployment-oriented evaluation, we report not only detection accuracy but localization robustness and computational cost through precision, recall, F1-score, mAP@50, mAP@50:95, inference time, FPS, model size, number of parameters, GFLOPs and training time. A single accuracy-efficiency score is introduced to help practitioners choose a model that fits real operational constraints. In the experiments, YOLOv8s delivers the strongest localization, reaching 0.9824 mAP@50 and 0.8825 mAP@50:95, together with the highest recall (0.9647). YOLO11s leads on precision and F1 -score, with values of 0.9814 and 0.9593, respectively. In contrast YOLOv8n offers the most suitable deployment profile, with 6.09ms inference time, 164.08 FPS and the best overall accuracy--efficiency score. Together, these findings indicate that real-time oil spill monitoring should be assessed jointly on accuracy, strict localization and compute cost, rather than ranked using mAP@50 alone.
Mohamed Mahmoud Ain Dhib, M. Lachgar, Mohamedou Cheikh Tourad et al.· EPJ Web of Conferences· 0 citations
A deep learning-based forensic framework for real-time detection of suspicious human activity in CCTV videos, trained without relying on any external sensors is proposed, and incorporates anonymization of personal identities and local edge-based processing to prevent raw data exposure.
Qazi Mazhar Ul Haq, Muhammad Imran, M. Waqas et al.· Arab Journal of Forensic Sci...· 0 citations
This work proposes a detection framework called distillation alignment YOLO (DA-YOLO) for PPE detection and introduces a teacher–student distillation framework with consistency constraints across predictions and high-order features extracted from baseline that enables the student model to achieve strong generalization while maintaining low computational cost.
Chonghua Zhou, Ruixuan Zhang, Yi-Xin Fu et al.· Multimedia Systems· 0 citations
Conventional video surveillance based on pixel-level deep-learning models is resource hungry, processes gigabytes of video material, and retains biometric identifying data. This paper describes a lightweight, privacy-sensitive alternative that uses skeletal pose estimation to replace pixel-based processing. We only process 33 coordinates for body joints using MediaPipe BlazePose, compressing the data by 2000× and discarding all visual identity data at the outset. The innovation is a hybrid dual- layer classifier that combines a Random Forest classifier trained on spatiotemporal features and deterministic rules from physics (velocity thresholding and kinematic plausibility), these rules help handle ambiguous cases when the Random Forest classifier is uncertain. Testing on 1145 frames of normal, suspicious, and dangerous activities achieves an overall accuracy of 87.42% with 0.52 precision and 0.89 recall for dangerous activities. The system has under 15 ms inference time on standard laptop CPUs and under 35 ms on Raspberry Pi 4, supporting real-time edge computing without GPUs. The framework reduces computation compared to CNN-based approaches and can run on $50 edge devices. Experiments show that skeletal geometric properties are sufficient for behavior classification without relying on appearance-based biometric features.
S. L. Jany Shabu, P. Asha, P.Asmitha Priyaa et al.· 2026 7th International Confe...· 0 citations
Background: Human-elephant conflict poses a significant threat to both wildlife conservation and rural livelihoods, particularly in regions bordering forest reserves. Traditional observation methods are often time-consuming, error-prone and limited under challenging environmental conditions, highlighting the need for automated detection systems. Methods: This study employs the YOLOv8 deep learning framework for elephant detection in thermal imagery. A publicly available thermal elephant dataset from Roboflow was preprocessed to remove low-quality and corrupted images. The dataset included diverse elephant postures, distances and environmental conditions with YOLO-formatted bounding box annotations. YOLOv8 was fine-tuned via transfer learning, utilizing multi-scale detection to localize elephants accurately across varying sizes and thermal scenarios. Result: The model demonstrated stable convergence, with box, segmentation, classification and distribution focal losses decreasing consistently over 50 training epochs. Detection performance was high, achieving a precision of 0.964, recall of 0.903, mAP@50 of 0.937 and mAP@50-95 of 0.676. Qualitative evaluation confirmed accurate localization under low contrast, motion blur and occlusion. These results indicate that YOLOv8, combined with rigorous dataset preprocessing, provides reliable real-time elephant detection for forest surveillance and early warning systems.
Seng-phil Hong· Indian Journal of Agricultur...· 0 citations
The YOLOv12 network is adopted as the baseline model and the ADown module is introduced to improve downsampling efficiency while maintaining lightweight performance, and the BN-CGLU is incorporated into the A2C2f module to enhance the model’s nonlinear representation capability.
Aobo Yue, Puchun Chen, Yan Yang· Journal of Real-Time Image P...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
