变电站环境下多目标多姿态行人语义分割检测

Semantic Segmentation-based Multi-target and Multi-posture Pedestrian Detection in Substation Environment

通过跳跃连接方式构建U-Net++深度学习模型,增加二进制交叉标记和骰子系数的组合作为损失函数,然后改进激活函数,并将Res2Net多尺度骨干架构作为特征提取网络,以获得更强的多尺度特征提取能力,进而对变电站环境下多目标多姿态行人进行语义分割检测实验。实验结果表明,对于变电站内行人目标极小、行人密集、行人多姿态非直立、行人多尺度并存等不同情况,所提算法都可准确检测并分割出行人目标,F1-Score指标可达0.978,检测效率可达0.035 s/张,优于其他两种经典方法。

Through the jump connection method,the U-Net++ deep learning model is constructed.The combination of binary cross-marking and dice coefficients is added as the loss function,then the activation function is improved,and the Res2Net multi-scale backbone architecture is used as the feature extraction network to obtain stronger multi-scale feature extraction ability.Then the semantic segmentation-based detection experiment is carried out on multi-target and multi-posture pedestrians in substation environment.The experimental results show that the pedestrian targets can be detected and segmented accurately under the conditions of very small pedestrian targets,dense pedestrians,non-upright pedestrian gestures and the coexistence of pedestrian multi-scales,with F1-Score index reaching 0.978 and detection efficiency of 0.035 s per image,which are superior to the two selected classical methods.