基于卷积神经网络的行人检测方法

A pedestrian detection method based on convolutional neural network

针对行人检测算法未能充分利用行人的特征信息,导致对行人的检测效果不佳问题,本文对无锚框的行人检测网络模型CSP进行了相应改进,提出了一种基于卷积神经网络的行人检测算法。首先,将原主干网络由ResNet-50加深为ResNet-101,然后引入卷积块注意力模块(CBAM)来提高原网络对小尺度行人中心点的特征表达,加入基于分数融合公式的分类器模块来进一步提高被遮挡行人的置信度,最终得到AS-CSP算法。该算法可以进一步提高对小尺度行人以及遮挡行人的检测效果。实验采用的数据集是CityPersons数据集,并在通用行人、小尺度行人以及遮挡行人等不同场景下进行对比实验,验证新算法的有效性。实验结果表明,本文提出的AS-CSP算法在通用行人、小尺度行人以及遮挡行人场景上的检测效果相比于原算法都得到了提升。

In view of the fact that the current pedestrian detection algorithm does not make full use of the pedestrian feature information, resulting in poor detection effect of pedestrians, this paper improves the pedestrian detection network model CSP and proposes a pedestrian detection algorithm based on convolutional neural network. Firstly, the original backbone network is deepened from ResNet-50 to ResNet-101, and then the convolution block attention module( CBAM) is introduced to improve the feature representation of small-scale pedestrian center points in the original network. Then, the classifier module based on the score fusion formula is added to further improve the confidence level of obscured pedestrians. Finally, the AS-CSP algorithm is obtained, which can further improve the detection effect of small-scale and obscured pedestrians. The experiment uses CityPersons data sets to carry out comparative experiments in different scenes(e.g. common pedestrians, small-scale pedestrians, obscured pedestrians,etc.) to verify effectiveness of the new algorithm. The experimental results show that the AS-CSP algorithm has higher detection effect of general pedestrians, small-scale pedestrians and occluded pedestrians than the original algorithm.