Efficient vehicle recognition in non-homogeneous and congested traffic situations is still a major challenge to smart transportation systems, especially in developing countries like Bangladesh, where the classes of vehicles are extremely different in size, looks, and movement patterns. The paper introduces a detailed analysis of the YOLOv8-s, YOLOv8-m and YOLOv8-l variants and suggests a more powerful architecture, which is named YOLOv8-L-CBAM, to enhance the detection stability in occlusion, changes in scale and visual complexity. The proposed model incorporates Convolutional Block Attention Module (CBAM) in the YOLOv8-l backbone to further optimise the features by highlighting channel-wise and spatially salient data. A variety of vehicle dataset that depicts the real-life Bangladeshi traffic were used as experimental, which includes cars, buses, trucks, motorcycles, rickshaws, CNGs, and easy bikes. Vastly compared based on the precision and recall and F1-score and accuracy and confusion-matrix-derived measurements, along with a larger engine, YOLOv8-l demonstrates better performance than smaller engines, and the proposed YOLOv8-L-CBAM is the best in terms of overall performance, with an average accuracy of 98.8% and higher mAP50-95. Specifically, significant positive changes are realised in the case of visually ambiguous and small vehicle categories like rickshaws and easy bikes. A scalable web-based deployment framework is built to verify real-world applicability and provide low-latency inference and human-in-the-loop verification. The findings confirm that attention-based feature refinement together with a powerful deployment pipeline is an efficient and feasible solution to real-time traffic monitoring and intelligent transportation applications.

Vehicle Detection,Convolutional Block Attention Module,Attention-Guided Deep Learning,Intelligent Transportation Systems,Real-Time Traffic Analytics