Abstract:

In order to improve the accuracy and speed of people counting in video monitoring and solve the problem of high occlusion caused by clothing blocking in traditional human body detection, an improved lightweight head detection MKYOLOv3-tiny is proposed. This method is an improvement on YOLOv3-tiny network. For improving the detection accuracy, multi-scale fusion is carried out on the head features of lower layer to realize classification prediction and position regression of different convolution layers. For improving anchors accuracy, K-means clustering algorithm is adopted to reduce initial anchors specification, according to the characteristics of small human heads and the idea of effective receptive field. Experimental results show that the detection accuracy of the improved model on Brainwash dataset increases by 3.21%, and the miss detection rate reduces by 8.7.

Key words: head detection, multi-scale fusion, K-means, effective receptive field, crowd counting

摘要：

为了提高视频监控中人数统计的精度和速度，解决传统人体检测由于衣物身体阻挡而造成的高遮挡问题。提出一种改进的轻量人头检测方法MKYOLOv3-tiny。该方法是对YOLOv3-tiny进行改进，针对低层的人头特征进行多尺度融合，实现不同卷积层的分类预测与位置回归，提升检测的精度；针对人头较小的特点，结合有效感受野的思想，K-means聚类减小初始候选框的规格，提升候选框的精度。实验结果表明，改进后的模型在Brainwash密集人头检测数据集上与原方法相比，在精度上提升了3.21%，漏检率降低了8.7%。

关键词: 人头检测, 多尺度融合, K-means, 有效感受野, 密集人数统计