Video Stabilization with YOLOv8 Object Detection for Surveillance Application

This research study intends to stabilize videos using deep learning without requiring an extensive dataset of paired shaky and stable videos, a challenge previously faced by traditional methods. By integrating concepts from traditional stabilization methods with deep neural networks, our method, DUT, addresses real-world issues with video stabilization. DUT comprises two stages: motion estimation and trajectory smoothing, yielding superior results compared to conventional methods. Additionally, this study focuses on detecting counterfeit car logos on roadsides using YOLOv8, an advanced object detection model. Leveraging videos captured from roadside cameras, our approach involves preprocessing steps such as video segmentation, frame selection, resizing, and normalization, followed by optional data augmentation. Bounding boxes are delineated around car logos, with class labels assigned for training. Through transfer learning, the YOLOv8 model is fine-tuned on the dataset, integrating single-shot detectors (SSDs) to accurately identify fake car logos in real-world environments. Rigorous experimentation and evaluation demonstrate significant enhancements in accuracy, offering an efficient solution to combat counterfeit car logos on roadsides.