Abstract:

In crowded places, security check is an important means to ensure public safety. Artificial problem for security in the event of emergencies or peak passenger flow, the efficiency and accuracy of screening is susceptible and there are security risks, an improved target detection algorithm Dense-YOLO is proposed based on YOLO algorithm. Characterized by drawing a dense network convergence to improve network structure, target block dimension is clustered by using a modified K-means algorithm in an abnormal image data set. Calculation efficiency is lifted by convolution and batch normalization integration, and increasing model robustness is increased to different sizes by multi-scale training. Experimental results show that the improved Dense-YOLO algorithm improves the detection of small targets, and detects the suspicious objects in the security inspection. The mAP reaches 91.68% and the detection speed increases to 59?f/s. The improved algorithm effectively improves the efficiency and accuracy of security checking, which eliminates security risks to a certain extent.

Key words: suspicious object detection, YOLO algorithm, Dense-YOLO algorithm, K-means algorithm, multi-scale training

摘要：

在人口密集场所中，安检是保证公共安全的重要手段。针对人工安检在遇到客流高峰或突发情况时，安检的效率和准确率易受到影响且存在安全隐患的问题，基于YOLO算法，提出了一种改进的Dense-YOLO目标检测算法。通过借鉴稠密网络中特征融合方式改进网络结构；采用改进的K-means算法对自制异常图像数据集进行目标框维度聚类；将卷积层中的卷积和批量归一化进行整合，提升计算效率；采用多尺度训练方式，增强模型对不同尺寸的鲁棒性。实验结果表明，利用改进后的Dense-YOLO算法提升了对小目标的检测，针对安检中可疑物进行检测，mAP达到了91.68%，检测速度提高到59?f/s。改进后的算法有效提升了安检的效率和准确率，一定程度上消除安全隐患。

关键词: 可疑物检测, YOLO算法, Dense-YOLO算法, K-means算法, 多尺度训练