In response to the problems of insufficient detection accuracy in complex scenarios, poor real-time performance in resource constrained environments, and weak vehicle traceability capabilities in the supervision of illegal tobacco transportation, this study constructs a "detection recognition deployment" multi-model collaborative optimization system: improving RT-DETR (RT-DETR-R18-P2) to solve high-precision detection in complex scenarios, improving ResNet34 (ResNet34-SE) to achieve vehicle attribute traceability, and adapting YOLO series models (YOLOv5/v8) to resource constrained deployment, achieving complementary advantages through "functional division-target collaboration-dynamic scheduling". Verified by 3580 vehicle images from the tobacco checkpoint in Qujing, Yunnan, the mAP50 of the collaborative system in complex scenes reached 92.8%, the inference speed of edge devices was 150 FPS, and the accuracy of attribute recognition was 93.8%. This formed an intelligent supervision scheme with "full scene coverage and full resource adaptation", providing technical support for the prevention and control of illegal tobacco transportation.