Abstract:

Safety helmets worn by workers is an important part of safety construction, to protect workers’ lives and overcome the defect of manual inspection, a new helmet detection method based on improved Single Shot MultiBox Detector（SSD） is proposed. In consideration of objects’ distribution in helmet dataset is imbalanced and objects are generally small, a branch network for feature fusion is added to SSD, it can enhance the shallow feature maps’ semantics and the mAP（mean Average Precision） of SSD300 is increased by 2.3 percentage points, and the real-time detection rate is reduced by only 1.3 frame/s, reaches 39.6 frame/s. In order to make SSD’s prior boxes match with the effective receptive field, the default box setting method is improved and the size of prior boxes is adjusted indirectly by introducing a variable parameter. The mAP of SSD300 and SSD512 reach 74.6% and 82.5% respectively. The experimental results show that the improved SSD model has excellent accuracy and good real-time performance and basically meets the requirements of practical application.

Key words: deep learning, computer vision, Single Shot MultiBox Detector（SSD）, helmet detection, feature fusion, small object

摘要：

施工人员佩戴安全帽是安全生产的重要一环，为保障工人生命安全，同时克服传统人工巡检费时费力的缺点，提出了一种基于Single Shot MultiBox Detector（SSD）改进的安全帽检测新方法。针对安全帽数据集内目标尺度偏小，尺度分布不均衡，对SSD模型结构进行改进，添加用以特征融合的分支网络，增强浅层特征图语义，引入该网络后SSD300的mAP-50（mean Average Precision）相应提升2.3个百分点，且SSD300实时检测速率仅降低1.3 frame/s，达到39.6 frame/s。为使SSD模型的先验框与有效感受野匹配，对SSD默认框设置方法进行改进，引入可变参数间接调节先验框大小，改进后的SSD300与SSD512的mAP分别达到74.6%与82.5%。安全帽数据集测试结果表明，改进后的SSD模型对安全帽佩戴检测具有优秀的准确性与良好的实时性，基本满足实际应用需求。

关键词: 深度学习, 计算机视觉, SSD, 安全帽检测, 特征融合, 小目标