铁路周界入侵目标多尺度特征感知算法

Multi-Scale Feature Perception Algorithm for Railway Perimeter Intrusion Object

准确识别侵入周界范围内的人和大型牲畜是铁路周界入侵视频智能分析技术的重点内容,对保障铁路安全运营具有重要意义。基于现有目标检测算法难以处理铁路监控场景中入侵目标呈现显著尺度变化的状况,提出一种多输入双输出神经网络(Multiple Input Double Output Network,MIDO-Net)和基于自适应特征加权融合的目标多尺度特征感知算法。首先,通过MIDO-Net多层级联的多输入和双输出网络结构,提取图像目标更丰富的多尺度特征信息;其次,依据骨干网络多阶段的特点,先将多级特征上采样至统一分辨率,再利用注意力模块和自适应参数对多级特征进行加权;然后,将特征输入到检测头中完成铁路周界入侵的识别;最后,通过视觉目标类别(Visual Object Classes,VOC)公共数据集和制作的多场景、多尺度铁路异物侵限数据集,对算法进行验证。结果表明:提出的多尺度特征感知算法在VOC公共数据集中的检测精确率达83.3%,在多场景、多尺度铁路异物侵限数据集中的检测精确率达91.1%,平均召回率达56.2%,均优于当前广泛使用的各种特征提取骨干网络;算法检测速率为45帧·s^(-1),优于同类型的骨干网络,且能满足铁路场景的行人实时监测需求。

Accurately identifying human and large livestock intruding within the perimeter is a key focus of intelligent video analysis technology for railway perimeter intrusion.It is of great significance for ensuring railway safety operations.However,existing object detection algorithms struggle to handle significant scale variations of intrusion objects in railway monitoring scenarios.Therefore,a Multiple Input Double Output Network(MIDO-Net)and a multi-scale feature perception algorithm based on adaptive weighted fusion are proposed.Firstly,the MIDO-Net extracts richer multi-scale feature information of image objects through its multi-level cascaded multiple input and double output network structure.Secondly,based on the multi-stage characteristics of the backbone network,the multi-level features are sampled up to unified resolution and then weighted using attention modules and adaptive parameters.Then,the features are input into the detection head to complete the recognition of railway perimeter intrusion.Finally,the algorithm is validated using the Visual Object Classes(VOC)public dataset and a self-made dataset of railway foreign object intrusion in multiple scenes and scales.The results show that the proposed multi-scale feature perception algorithm achieves a detection accuracy of 83.3%in the VOC public dataset and 91.1%in the dataset of railway foreign object intrusion in multiple scenes and scales.The average recall rate is 56.2%,which is superior to various widely used feature extraction backbone networks.The algorithm detection rate is 45 frames per second(fps),surpassing similar backbone networks,and can meet the requirements for pedestrian real-time monitoring in railway scenarios.