人群分布不均、遮挡和背景干扰等问题使得人群计数成为了一项复杂且具有挑战性的任务.针对这些问题,提出了一种多尺度特征融合的位置关注网络(Position-Aware Network based on Multi-Scale Feature Fusion,MSF-PANet).首先,设计了一种多尺度特征融合模块,以在不同感受野下提取并融合人群密度图的多尺度特征,同时提取出前景信息,来应对人群计数中的遮挡和背景干扰问题;然后,通过位置注意力分配网络提高模型对人群区域的关注度,有效地应对人群分布不均的问题;最后,为了辅助模型训练,减小背景噪声带来的干扰,引入了一种结构交叉损失用于强化模型对人群结构的学习.实验结果表明:MSF-PANet在Shanghai Tech Part A、Shanghai Tech Part B、UCF-QNRF和UCF_CC_50 上平均绝对误差分别为 59.5、7.8、103、182.7,均方误差分别为 96.7、13.6、177、237.7,验证了所提模块在提高人群计数准确率上的有效性.
Research on crowd counting based on multi-scale feature fusion and position-aware network
The uneven distribution of crowds,occlusion,and background interference have made crowd counting a complex and challenging task.In response to these issues,a Position-Aware Network based on Multi-Scale Feature Fusion(MSF-PANet)is proposed.Firstly,a multi-scale feature fusion module is designed to extract and integrate multi-scale features of crowd density maps at different receptive fields,while capturing foreground information to address occlusion and background interference issues.Secondly,a position attention network is employed to enhance the model's focus on crowd regions,effectively dealing with the problem of uneven crowd distribution.Lastly,a structural cross loss is introduced to assist model training,reducing the impact of background noise and strengthening the learning of crowd structure.Experimental results demonstrate the effectiveness of MSF-PANet,with average absolute errors of 59.5,7.8,103,and 182.7,and mean squared errors of 96.7,13.6,177,and 237.7 on Shanghai Tech Part A,Shanghai Tech Part B,UCF-QNRF,and UCF_CC_50 datasets,respectively,validating the improvement in crowd counting accuracy achieved by the pro-posed module.
crowd countingattention mechanismmulti-scale featurebackground segmentationcrowd density estimation
万方数据
维普
