改进YOLOv4-tiny网络的狭小空间目标检测方法

Narrow Space Object Detection Method by Improved YOLOv4-tiny Network

针对狭小空间中目标相互遮挡导致轻型检测网络存在大量漏检、分类错误等问题,基于YOLOv4-tiny提出一种自适应非极大抑制(adaptive non-maximum suppression,A-NMS)的多尺度检测方法。在骨干网络引入大尺度特征图优化策略和金字塔池化模型,增强遮挡目标显著区域特征;设计内嵌空间注意力的双路金字塔特征融合网络,提升浅层细节特征与高级语义信息的融合能力;提出区域目标密度与边界框中心距离因子相关联的动态NMS阈值设定方法,并在后处理阶段代替传统IoU-NMS算法,进一步减少漏检。实验结果表明,与YOLOv4-tiny算法相比,改进算法在公开数据集PASCAL VOC07+12和自制数据集上mAP值分别提高2.84个百分点和3.06个百分点,FPS保持在87.9,对遮挡目标的检测能力显著提升,满足移动端对狭小复杂场景实时检测的需求。

Aiming at the problems of a large number of missed detections and classification errors in the light detection network due to the mutual occlusion of objects in narrow spaces, an adaptive non-maximum suppression(A-NMS)multiscale detection method based on YOLOv4-tiny network is proposed. A large-scale feature map optimization approach and a pyramid pooling model are incorporated into the backbone network to enhance the major regional features of hidden objects. To improve the fusing of shallow detail data with high-level semantic information, a two-way pyramidal feature fusion network with embedded spatial attention is designed. A dynamic NMS threshold setting method that correlates the regional object density with the distance factor of the center of the bounding box is proposed and replaces the traditional IoU-NMS algorithm in the post-processing stage to further reduce the missed detection. The experimental results show that compared with the YOLOv4-tiny algorithm, the improved algorithm improves the mAP value by 2.84 percentage points and 3.06 percentage points on the public dataset PASCAL VOC07 + 12 and the self-made dataset, respectively,while the FPS remains at 87.9, the detection capability of occluded objects is significantly improved to meet the demand of mobile terminals for real-time detection of narrow and complex scenes.