基于改进YOLO V3算法的室内人数统计模型

Indoor People Counting Model Based on Improved YOLO V3 Algorithm

摘要: 基于机器学习与深度学习的目标检测方法被广泛应用于人数统计，然而实际检测区域往往存在人群相互遮挡，或光照不均匀等情况时，人数统计仍然面临很大挑战。为此，提出了一种改进的YOLO V3模型，使其更好的适应室内人群的人数统计。首先自建并丰富了数据集，增加了训练数据的多样性，并通过K-means算法重新聚类锚框；其次，提出了F-YOLO V3模型，该模型中增加104 × 104尺寸的特征图输出并取消13 × 13尺寸特征图的输出；将原网络每一层上采样后的特征图继续上采样，得到的特征图与原网络相应尺寸的特征图进行拼接；并将输出层前的5个卷积变成了1个卷积和2个残差单元，提取更多特征信息，增强对模糊或者较小目标检测能力；最后增加一个ADIOU Loss分支衡量检测框的定位准确度。实验结果表明，F-YOLO V3模型具有更高的召回率和平均精度，室内场景下的人员统计性能得到显著提升。

Abstract: Object detection methods based on machine learning and deep learning model are widely used in people counting. However, when there are too many objects in the same area, people will be oc-cluded, or people in the video are not easy to find in the dark, people counting is still a big challenge. Therefore, an improved YOLO V3 model is proposed to better adapt to the number of indoor crowd statistics in classrooms. Firstly, the data set was self-built and enriched to increase the diversity of training data, and the anchor boxes were re-clustered by K-means algorithm. Secondly, the YOLO V3 feature extraction network and multi-dimension detection algorithm were improved, and the F-YOLO V3 model was proposed. In this model, the output of 104 × 104 feature map was added and the output of 13 × 13 feature map was canceled. The sampled feature images of each layer of the original network are continued to be sampled, and the obtained feature images are spliced with the corresponding size feature images of the original network. The 5 convolutions in front of the output layer are changed into 1 convolution and 2 residual units to extract more feature information and enhance the detection ability of fuzzy or small targets. Add an ADIOU Loss branch to measure the positioning accuracy of the detection box; Finally, the real-time number of people in the output screen is counted. The experimental results show that the improved YOLO V3 algorithm has higher recall rate and average precision, and the performance of personnel statistics in indoor scenes is significantly improved.