Dense crowd counting method of multi-scale convolutional neural network

Abstract

The invention discloses a dense crowd counting method of a multi-scale convolutional neural network, a network structure adopting an SPMsCNN algorithm adopts a multi-scale convolutional neural networkalgorithm based on accumulation pooling, a multi-scale module in the algorithm is provided with a plurality of filters with different sizes, more character feature information is extracted, and the information is reserved through feature splicing. Meanwhile, accumulation pooling is introduced, a traditional pooling layer is improved under the condition that parameters and hyper-parameters are notintroduced, the characteristic that the maximum pooling can guarantee the scale invariance of a network model is utilized, the pooling layers with small filters are overlaid, the scale invariance isguaranteed in a larger range, fine-grained information is reserved, and meanwhile, the pooling performance of the network model is improved. The number of network structure layers of the SPMsCNN algorithm is not large, and the phenomenon of gradient disappearance or gradient explosion cannot be caused.

Description

【Technical field】 [0001] The invention relates to big data statistics and collection algorithm technology, in particular to a multi-scale convolutional neural network dense crowd counting method. 【Background technique】 [0002] The methods of counting the number of dense crowds can be roughly divided into three categories: detection-based methods, regression-based methods, and density estimation-based methods. The detection method, as the name implies, is to detect people. As a detection standard, except The overall characteristics of the person, as well as the characteristics of the unique parts of the person, such as the head, shoulders, ears, etc., in order to effectively reduce the impact of the occlusion problem on the experiment, many detection-based methods are based on the detection of human body parts; The detection-based method first needs to preprocess the picture, then extract the foreground of the picture, and then use the detector to judge which is the person i...

AI Technical Summary

Problems solved by technology

Looking at all network structures, it is not difficult to find that deep convolutional neural networks and multi-column convolutional neural networks are the most used. Although deep convolutional neural networks and multi-column convolutional neural networks perform well in dense crowd scenes, but The deep convolutional neural network is easy to cause overfitting, and because the depth of the network structure is too deep, there may be a problem of gradient disappearance, which also increases the difficulty of training the network model. It is difficult to train multi-column convolutional neural networks. First, each column needs to be pre-trained, and then the whole is trained, and the network structure with shallower depth and fewer columns performs poorly in feature extraction, and is easily affected by the multi-scale problem of the image.

Images