【摘要】 在特种设备操作考试中，为提升评判的客观性与监考的智能化水平，无人化监考系统正逐步应用于叉车实操场景。针对该类系统对车轮位置与车道线精确识别的需求，提出一种融合目标检测与语义分割的视觉识别方法，通过分阶段处理策略，结合目标检测快速定位能力和语义分割精细识别优势，实现车轮位置与车道线的高精度识别。该方法采用YOLOv8n模型实现叉车位置的快速检测，以保障系统实时响应能力。同时引入优化的DeepLabv3+模型进行车道线分割，其主干网络采用MobileNetv3模块，减少模型参数量与计算开销，并引入卷积块注意力模块(CBAM)以增强对关键区域的感知能力。在分割头中融合空洞非对称卷积块-空洞空间金字塔池化(AACB-ASPP)与多级特征融合模块(MFFM),提升对多尺度目标与细节的表达能力。实验结果表明，该方法在验证集上达到84.29%的平均交并比(mIoU)和90.45%的平均像素准确率(mPA),较基线模型分别提升2.09%和3.10%。所提方法具备部署灵活、维护简便的优势，可为叉车考试中的无人化智能监考系统提供高效、可靠的视觉识别支持，具有广泛应用潜力。更多还原

【Abstract】 In special equipment operation examinations, unmanned supervision systems are being progressively applied to forklift practical operation scenarios to enhance the objectivity of evaluation and the intelligence level of supervision. To address the requirement for precise identification of wheel positions and lane lines in such systems, a visual recognition method integrating object detection and semantic segmentation is proposed. By employing a phased processing strategy that combines the rapid localization capabilities of object detection with the fine-grained recognition advantages of semantic segmentation, this approach achieves high-precision identification of wheel positions and lane lines. The method employs YOLOv8n to achieve rapid detection of forklift positions, ensuring real-time response capability of the system. Simultaneously, an optimized DeepLabv3+ network is introduced for lane line segmentation, with MobileNetv3 as the backbone network to reduce model parameters and computational overhead, and the convolutional block attention module(CBAM) is incorporated to enhance the perception capability for key regions. In the segmentation head, atrous asymmetric convolution block-atrous spatial pyramid pooling(AACB-ASPP) and multi-level feature fusion module(MFFM) are integrated to enhance the representation capability for multi-scale targets and details. Experimental results demonstrate that the method achieves 84.29% mean intersection over union(mIoU) and 90.45% mean pixel accuracy(mPA) on the validation set, representing improvements of 2.09% and 3.10% respectively compared to the baseline model. The proposed method possesses advantages of flexible deployment and convenient maintenance, and can provide efficient and reliable visual recognition support for unmanned intelligent supervision systems in forklift examinations, demonstrating broad application potential.更多还原