<![CDATA[Real-time processing and accuracy of surveillance systems in a real-world environment are very critical to the understanding of the scene. Current deep learning methods, however, have failed to sufficiently learn the global context and relations between objects from high resolution surveillance information. In this study, a hybrid deep learning framework is proposed, which incorporates the merits of Vision Transformers, Graph Neural Networks, and attention-based fusion to tackle these challenges. The novelty of the research is that they design a single architecture combining global features extraction, and relational reasoning to better interpret the surveillance scene. The proposed model has an improved representation of features, which includes complementary mechanisms in a single model. In terms of method, Vision Transformers are used to use global spatial features from surveillance frames. Relationships between the detected objects are described in a structured graph using the Graph Neural Networks. In complex scenes, these features are adaptively integrated together through an attention guided fusion scheme, which improves the classification and understanding of complex scenes. Experimental results on UA-DETRAC dataset show that the proposed framework achieves higher accuracy, precision, recall, F1 score, mAP and AUC (0.962%, 0.955%, 0.958%, 0.956%, 0.960% and 0.97 respectively). Furthermore, the model's performance is 78 fps making it a compatible model for real-time surveillance systems. Analysis of these results shows that the method is robust and improved consistently compared with the baseline methods in all performance measure. Last but not least, the proposed framework is very efficient and accurate, and is able to achieve a good balance between accuracy and computation efficiency in real-time intelligent surveillance. The results prove its potential application for surveillance, traffic monitoring and security critical environments in smart cities.]]>
