改进YOLOv3的实时性视频安全帽佩戴检测算法

Improved Algorithm of YOLOv3 for Real-Time Helmet Wear Detection in Videos

生产安全事故频繁发生,佩戴安全帽可以有效避免人员伤亡。提出一种在自然场景下是否佩戴安全帽的检测方法,对YOLOv3算法进行改进。使用ResNeXt50作为新的特征提取网络结合CSP Net中的梯度分流截断思想提高算法的特征表达能力,融合SPP Net和PA Net提高算法特征融合的质量,使用CIoU优化损失函数。并且结合DeepSort目标跟踪技术提高算法实时性。在SHMD安全帽数据集上进行验证,算法平均准确率达到96.80%,较原算法提升12%,且运行速率达到32帧/s。在多种场景下,都具有较好的检测效果。

Production safety accidents happen frequently,wearing safety helmets can effectively avoid casualties.This paper presents a method to detect whether a helmet is worn in natural scenes by improve the YOLOv3 algorithm.ResNeXt50 is used as a new feature extraction network and combine the idea of gradient shunt truncation in CSP Net to improve the feature extraction capability of the algorithm,SPP Net and PA Net are fused to improve the quality of feature fusion of the algorithm,and CIoU is used to optimize the loss function.DeepSort multi-object tracking technology is combined to improve the real-time performance of the algorithm.The algorithm was verified on the SHMD helmet dataset with an average accuracy of 96.80%,an improvement of 12%over the baseline,and operating rate of 32fps.In many scenarios,it has a good detection effect.