一种基于Mobilenetv3的行人检测算法研究

Research on pedestrian detection algorithm based on Mobilenetv3

针对实时的行人检测算法要求模型具有轻量型和良好的鲁棒性,文中提出一种基于Mobilenetv3的行人检测算法。该算法首先采用Mobilenetv3作为模型的主干特征提取网络;然后通过深度可分离卷积替换PANet中的普通卷积,减少网络的复杂度;最后引入注意力机制SE和ECA关注网络中重要的通道信息,加强模型的特征融合能力。实验结果表明:与YOLOv4算法相比,基于Mobilenetv3的行人检测算法模型体积缩小78.03%,参数量也降低82.44%;且在实验数据集和INRIA数据集上,所提算法的平均精度(AP)分别提升3.98%和1.10%,检测速率分别提升8.08 f/s和7.89 f/s,检测时间也显著缩短,具有良好的检测性能。

As the real-time pedestrian detection algorithm requires the model to be lightweight and good robust,a pedestrian detection algorithm based on Mobiletv3 is proposed.In the algorithm,the Mobiletv3 is used as the backbone feature extraction network of the model,and then the ordinary convolution in PANet is replaced by means of the deep separable convolution to reduce the complexity of the network.The attention mechanism SE and ECA are introduced to pay attention to the important channel information in the network to strengthen the feature fusion ability of the model.The experimental results show that,in comparison with YOLOv4 algorithm,the model volume of the pedestrian detection algorithm based on Mobilenetv3 is reduced by 78.03%,and its parameter quantity is also reduced by 82.44%.On the experimental dataset and INRIA dataset,the average accuracy(AP)of the proposed algorithm has been improved by 3.98%and 1.10%,respectively,and the detection rate has been improved by 8.08 f/s and 7.89 f/s,respectively.The detection time has also been shortened significantly,indicating good detection performance.