The demand for efficient real-time surveillance systems has increased significantly due to the growing need for enhanced security in public and private spaces. However, such systems face a significant challenge in achieving high accuracy of object detection while also ensuring efficient compression of the input images, particularly in bandwidth-limited setups. In this paper, we propose a new architecture that uses an Adaptive Query Vision Transformer (AQ-ViT) for enhanced object detection with a Novel Learned Image Compression Network (NLIC) for transmission and storage at runtime surveillance settings. The AQ-ViT employs a dynamic query mechanism with multi-scale attention fusion, allowing it to adaptively detect objects of different sizes and complexities across various surveillance scenarios. Meanwhile, NLIC adopts the hierarchical latent representations and adaptive bit allocation to effectively retain the important visual information in the source, while providing better compression performance. The proposed framework is evaluated through extensive experiments on standard benchmarks and real-world surveillance datasets to show that it outperforms state-of-the-art object detection and image compression approaches in terms of accuracy, compression ratio, and latency. This work unifies detection and compression in one architecture, providing a scalable and efficient approach for next-generation surveillance systems that maintain reliable performance in resource-constrained scenarios. The new framework increases the accuracy of visual threat detection to 7.74% as well as compressing the data to 2.70% which can be implemented in surveillance deployments based on IoT.