To address the challenges of simultaneously satisfying detection accuracy, edge real-time performance, low-power operation, and end-to-end business linkage in parking scenarios, this paper proposes an intelligent parking barrier system based on deep learning and multi-sensor fusion. The system adopts a three-layer collaborative architecture comprising an edge sensing node layer, a cloud business service layer, and a front-end management application layer. On the edge side, a Raspberry Pi 5 integrates a camera, infrared ranging sensor, MPU6050 attitude sensor, and LoRa module for parking-state sensing and local decision-making. At the algorithmic level, YOLOv3-tiny is structurally pruned for single-class detection, compressing model weights to approximately 33 MB. At the decision level, an asymmetric infrared-vision-inertial fusion state machine is designed, employing an "infrared trigger - visual confirmation - inertial fallback" mechanism to enhance robustness under nighttime, occlusion, and impact disturbances. Experimental results show that after over 5000 training iterations, [email protected] reaches 96.5%-98.2%. On Raspberry Pi 5, single-frame inference latency at 416x416 resolution is 600-850 ms, meeting polling requirements of 5 s (idle) and 10 s (occupied). Average power consumption decreases from 4.02 W to 1.02 W, achieving approximately 74% energy savings. Joint debugging tests further validate the solution's advantages in detection accuracy, response timeliness, energy efficiency, and engineering deployability.

翻译：为解决停车场景下同时满足检测精度、边缘实时性、低功耗运行与端到端业务联动等挑战，本文提出一种基于深度学习与多传感器融合的智能停车道闸系统。系统采用边缘感知节点层、云端业务服务层与前端管理应用层三层协同架构。在边缘侧，树莓派5集成摄像头、红外测距传感器、MPU6050姿态传感器及LoRa模块，实现停车状态感知与本地决策。算法层面，对YOLOv3-tiny进行结构剪枝以适配单类检测任务，模型权重压缩至约33 MB。决策层面，设计非对称红外-视觉-惯性融合状态机，采用“红外触发-视觉确认-惯性回退”机制增强夜间、遮挡及撞击干扰下的鲁棒性。实验结果表明，经5000次以上训练迭代后，[email protected]达到96.5%-98.2%。在树莓派5上，416×416分辨率下单帧推理时延为600-850 ms，满足5秒（空闲）与10秒（占用）轮询需求。平均功耗从4.02 W降至1.02 W，实现约74%节能。联合调试测试进一步验证了方案在检测精度、响应时效、能耗效率与工程可部署性方面的优势。