基于多层BP神经网络和无参数微调的人群计数方法

Crowd Counting Method Based on Multilayer BP Neural Networks and Non-parameter Tuning

针对大部分现有的人群计数方法被应用到新的场景时性能下降的问题,在多层BP神经网络框架下,提出一种具有无参数微调的人群计数方法。首先,从训练图像中裁切图像块,将获得的相似尺度的行人作为人群BP神经网络模型的输入;然后,BP神经网络模型通过学习预测密度图,得到了一个具有代表性的人群块;最后,为了处理新场景,对训练好的BP神经网络模型进行目标场景微调,可追求有相同属性的样本,包括候选块检索和局部块检索。实验数据集包括PETS2009数据集、UCSD数据集和UCF_CC_50数据集。这些场景的实验结果验证了提出方法的有效性。相比于全局回归计数法和密度估计计数法,提出的方法在平均绝对误差和均方误差方面均有较大优势,消除了场景间区别和前景分割的影响。

Because the performance of most existing crowd counting methods is decreased when they are applied to a new scene,a crowd counting method based on non-parameter tuning was proposed in the framework of multilayer BP neural network.Firstly,image blocks are cropped from the training images to obtain similar scale pedestrian as an input of crowd BP neural network model.Then,the predictive density map is learned by BP neural network model to obtain representative crowd blocks.Finally,in order to deal with the new scene,the target scene is adjusted on the trained BP neural network model,retrieving samples with the same attributes,which includes candidate block retrieval and local block retrieval.The data set includes PETS2009 data set,UCSD data set and UCF_CC_50 data set.The effectiveness of the proposed method is verified by the experimental results on these scenes.Compared with the global regression coun-ting method and density estimation counting method,the proposed method has advantages of average absolute error and mean square error,and overcomes the influences of the differences between the scenes and foreground segmentation.