ABSTRACT:

This project presents an advanced traffic object detection system designed for vehicle-mounted fisheye cameras. Traditional detection methods struggle with distortions and varying object scales introduced by fisheye imaging. To overcome these challenges, two novel deep learning architectures are employed: YOLO26-SPAC (Spatial Pyramid Attention and Context) and YOLO11-CFA (Cross-Frequency Attention). YOLO26-SPAC incorporates spatial pyramid attention and global context embedding, enabling multi-scale feature extraction and enhancing localization of traffic objects. YOLO11-CFA leverages cross-frequency attention and wavelet-based feature extraction to capture both low- and high-frequency components, improving robustness against geometric distortions. The system classifies traffic entities into five categories: Bike, Bus, Car, Pedestrian, and Truck. Data augmentation techniques such as mosaic, flipping, and scaling are applied to improve model generalization. The framework is developed with a front-end using HTML, CSS, and JavaScript, and a back-end using Python with the Flask framework. Evaluation metrics include precision, recall, F1-score, mAP@0.5, and mAP@0.5:0.95, with confusion matrices and validation predictions used to assess model performance. This approach demonstrates improved detection accuracy, better handling of distortions, and effective classification across multiple traffic object types.

Keywords: Traffic detection, Fisheye imaging, YOLO26-SPAC, YOLO11-CFA, Spatial Pyramid Attention, Cross-Frequency Attention, Wavelet features, Object classification, Deep learning, Feature fusion