Adaptive query allocation for dense object detection in deformable transformers

• The proposed model introduces a density map-guided query selection mechanism, enabling dynamic allocation of object queries to high-density regions. This ensures finer granularity in crowded areas without increasing computation in sparse regions. • Built upon the Deformable DETR framework, the model incorporates a flexible multi-resolution deformable attention module, improving the ability to detect overlapping and variably sized objects across spatial scales. • A novel density map generation module is integrated into the detection pipeline to provide prior knowledge of object concentration, effectively guiding downstream attention and query mechanisms. • The approach outperforms state-of-the-art detectors on challenging datasets (e.g., aerial imagery, autonomous driving) by addressing occlusion, scale variance, and close object proximity with adaptive architectural components. • Despite the adaptive and multi-scale mechanisms, the model maintains computational efficiency through end-to-end optimization, making it suitable for real-time applications in smart surveillance and robotics. In this paper, we propose a novel density-aware transformer-based model for object detection in densely crowded scenes such as aerial imagery and autonomous driving scenarios. Our approach builds upon the Deformable DETR framework, introducing key enhancements that address the challenges posed by overlapping and closely packed objects. First, we generate density maps for highlighting regions with high object concentration, providing valuable cues for the subsequent modules. Next, a dynamic object query selection mechanism utilizes these maps to adaptively allocate a suitable number of queries per region, ensuring finer detection granularity in cluttered areas without adding excess computational cost in sparser regions. Finally, a flexible multi-resolution deformable attention module attends to critical visual cues across varying spatial scales, allowing for more robust regression and classification of objects of different sizes. Through end-to-end training, our model automatically learns to capture local and global context, accurately identifying objects even when heavily occluded or partially visible. Experimental results on challenging benchmarks demonstrate that our method outperforms existing state-of-the-art detectors in terms of accuracy while achieving comparable computational efficiency. These improvements highlight the potential of our approach as a reliable solution for real-world applications requiring fast and precise detection in highly congested environments.