FlowDet: Overcoming Perspective and Scale Challenges in Real-Time End-to-End Traffic Detection

TL;DR

FlowDet addresses real-time traffic intersection detection by introducing geometry-aware sampling via a Geometric Deformable Unit and efficient multi-scale processing via Scale-Aware Attention, embedded in a Progressive Adaptive Feature Cascade. It also releases Intersection-Flow-5K, a challenging benchmark with dense, occluded, and small objects to evaluate detectors in realistic intersections. Empirical results show FlowDet achieving state-of-the-art accuracy while substantially reducing computations and increasing speed, outperforming RT-DETR and other baselines on Intersection-Flow-5K and COCO. The work demonstrates a practical path toward accurate, efficient, edge-deployable detectors for intelligent transportation systems.

Abstract

End-to-end object detectors offer a promising NMS-free paradigm for real-time applications, yet their high computational cost remains a significant barrier, particularly for complex scenarios like intersection traffic monitoring. To address this challenge, we propose FlowDet, a high-speed detector featuring a decoupled encoder optimization strategy applied to the DETR architecture. Specifically, FlowDet employs a novel Geometric Deformable Unit (GDU) for traffic-aware geometric modeling and a Scale-Aware Attention (SAA) module to maintain high representational power across extreme scale variations. To rigorously evaluate the model's performance in environments with severe occlusion and high object density, we collected the Intersection-Flow-5k dataset, a new challenging scene for this task. Evaluated on Intersection-Flow-5k, FlowDet establishes a new state-of-the-art. Compared to the strong RT-DETR baseline, it improves AP(test) by 1.5% and AP50(test) by 1.6%, while simultaneously reducing GFLOPs by 63.2% and increasing inference speed by 16.2%. Our work demonstrates a new path towards building highly efficient and accurate detectors for demanding, real-world perception systems. The Intersection-Flow-5k dataset is available at https://github.com/AstronZh/Intersection-Flow-5K.

Figures (5)

• Figure 1: Performance comparison on the Intersection-Flow-5K dataset. FlowDet establishes a new state-of-the-art Pareto curve, achieving superior accuracy at real-time speeds compared to advanced detectors like YOLOv9 YOLOv9 and RT-DETR DETRs.
• Figure 2: Overall architecture of FlowDet. The framework integrates PAFC with GDU in the backbone and SAA in the encoder for efficient traffic object detection.
• Figure 3: GDU within Adaptive Refinement Unit. The learned geometric-aware spatial offsets enable adaptive modeling for traffic objects with varying perspective characteristics.
• Figure 4: Intersection-Flow-5K captures a spectrum of critical challenges in traffic surveillance, including (a) sensor saturation from nighttime glare, (b) complex inter-object occlusion, (c) extreme scale variation with distant objects, and (d) baseline ideal conditions.
• Figure 5: Qualitative comparison showcasing FlowDet's superior attention mechanism and detection performance. Top row (Occlusion): In challenging scenarios with severe inter-object occlusion, FlowDet precisely focuses on visible vehicle parts, leading to robust detection. Bottom row (Small Objects): For distant, low-resolution targets, our model maintains a tight attention focus, demonstrating its resilience to extreme scale variations. FlowDet's refined geometric and scale-aware modeling results in more accurate and reliable detections in both cases.