A method and system for real-time detection of multimodal pedestrian images
By using the improved MAFD-DETR model and employing adaptive frequency-domain gated wavelet convolution and thermal gradient-guided windmill convolution modules, the problems of insufficient feature extraction and high false alarm rate in multimodal pedestrian detection under adverse lighting conditions are solved, achieving high-precision and real-time pedestrian detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUXI UNIV
- Filing Date
- 2026-03-23
- Publication Date
- 2026-05-26
AI Technical Summary
Existing multimodal pedestrian detection technologies struggle to adaptively suppress frequency domain noise under adverse lighting conditions, neglect edge gradient information in infrared images, and fail to balance high accuracy with real-time performance, resulting in inaccurate detection and high computational demands.
An improved MAFD-DETR model is adopted, which uses a dual-stream backbone network, including an adaptive frequency-domain gated wavelet convolution module and a thermal gradient-guided windmill convolution module, combined with a progressive semantic fusion module, to adaptively process visible light and infrared image features, thereby achieving multi-scale feature extraction and cross-modal feature fusion.
It significantly reduces the false alarm rate of small targets, improves the robustness and detection accuracy of the model under complex lighting conditions, meets real-time requirements, reduces the false alarm rate, and enables all-weather pedestrian detection.
Smart Images

Figure CN121884396B_ABST