针对轨道异物入侵检测算法对远距离小目标异物检测效果差、复杂天气场景下鲁棒性差等问题，提出一种基于LAM-YOLOv8n的全天候轨道异物入侵检测算法。首先，针对远距离小目标检测以及深层特征易丢失等问题，设计了一种轻量型多尺度高效的特征融合网络LPANet(Lightweight Path Aggregation Network),通过增加小目标检测层提高小目标检测精度，通过对浅层特征进行下采样与深层特征进行跳跃式连接融合更多的位置特征信息，通过减小低分辨率实现特征检测层轻量化；其次，结合可变形卷积DCNv2与1维通道卷积设计了一种具有残差结构的空间-通道型自适应注意力机制AS-LCAM(Adaptive Spatial-Lightweight Channel Attention Module),减少复杂背景下对于关键信息的干扰；最后，设计M-GhostC2f轻量化模块代替C2f模块，减少网络中存在的冗余特征，提高冗余特征映射过程中提取关键特征的能力。实验结果表明，所提出的LAM-YOLOv8n模型在实际轨道场景采集的数据集中mAP50指标达到89.2%,Params仅为1.9M,模型大小为4.2 MB,较YOLOv8n模型mAP50提高了4%,Params降低了36%,模型大小降低了31.1%,在轻量化的同时具有更高的检测效果。

The excitation amplitude and load also have great affection on the transmissibility performance. Aiming at the serious threat of intruding foreign objects within the train operation environment to the safety of trains, and the existing track foreign object intrusion detection algorithms have poor effect on the detection of small target foreign objects at a long distance, and poor robustness in complex weather scenarios, this paper proposes an all-weather track foreign object intrusion detection algorithm based on LAM-YOLOv8n. First, a Lightweight Path Aggregation Network(LPANet) is designed to address the problems of long-distance small object detection and the easy loss of deep features, which improves the small object detection accuracy by increasing the small object detection layer, fuses more positional feature information by downsampling the shallow features and jumping connection with the deep features, and realizes the lightweighting by reducing the low-resolution feature detection layer; Second, an adaptive spatial-lightweight channel attention module(AS-LCAM) with residual structure is designed by combining deformable convolutional DCNv2 with 1-dimensional channel convolution, which enhances the model's global context modeling capability and reduces the interference of critical information in complex contexts; Finally, the M-GhostC2f lightweight module is designed to replace the C2f module to reduce the redundant features present in the network and to improve the ability of extracting key features during the redundant feature mapping process. The experimental results show that the LAM-YOLOv8n model proposed in this paper has 89.2% mAP50, 1.9M Params, and 4.2MB model size in the dataset collected from the real track scene, which is 4% higher than the YOLOv8n model in terms of mAP50, 36% lower than Params, and 31.1% lower than the model size, and it is lighter while having a higher detection effect.