Bridging CNNs and transformers: a mamba-based detection framework for intelligent traffic scenarios

UAV object detection in intelligent transportation systems suffers from extreme scale variation, complex background confusion, and resource constraints of optical sensing applications. We present CY-DETR, a lightweight detection transformer framework for UAV aerial surveillance scenarios with optical sensors. We introduce three core innovative modules. First, CGNet blends Mamba state space modeling with grouped convolution, capable of modeling longer-range dependencies but with lower complexity. Second, SMRFPN merges soft nearest neighbor interpolation and adaptive feature enhancement modules that effectively preserve important spatial details for small object detection in optical images. Finally, EPG-optimized FSA dynamically implements sparse attention using content-aware importance assessment that processes high-resolution optical sensor data. Extensive experiments on standard UAV surveillance benchmarks verify that CY-DETR attains better detection accuracy with fewer parameters than previous work, in addition to observing inference times speed-up and significant parameter savings compared with baseline RT-DETR. In general the framework transfers better to other UAV surveillance scenarios and optical sensing conditions, performing on the trade-off between detection accuracy and computational efficiency and ultimately enabling real-time optical sensor-based object detection on low-power UAVs.