Research On The Scale Fusion Based Crowd Counting Algorithms

Massive crowd gathering will lead to stampede and other accidents.Therefore,it is necessary to monitor the crowd density in assembly scenes.Crowd counting is an important branch of computer vision,which is of great significance for urban safety.However,the individual freedom in reality leads to the random distribution of crowd and the huge variety of individual scale,which brings great challenges to the crowd counting algorithms.At present,there are some shortcomings in crowd counting algorithm: when generating the density map training labels,estimating the scale of heads according to the distance between the labeled points is not accurate;the regression neural network of crowd density map cannot effectively extract the multi-scale features in the crowd;there is noise when the crowd image is regressed to the density map.To solve the above problems,this paper studies the crowd counting algorithms and proposes a scale fusion based crowd counting algorithm SFBCC.It mainly includes the following innovations:(1)To solve the problem that the traditional crowd density map generation algorithm cannot accurately estimate the head size,a scale adaptive density map construction algorithm is designed.Multiple head detectors are used to obtain the head scale of the sparse crowd part,and radial basis function interpolation algorithm is used to complete these parts.As to the other dense area,the distance adaptive algorithm is adopted.A more accurate density map is generated by combining these algorithms.(2)In order to further improve the ability of network to extract multi-scale features,we draw lessons from Efficient Net and Efficient Det to design scale fusion based density map regression neural network.We use mobile flip bottleneck convolution module to design feature extraction skeleton network,and add dilated convolution module to further improve the ability of multi-scale feature extraction.Bi FPN is adapted to integrate feature maps with different scales.(3)In order to suppress the noise in non-crowd regions when generating density map,the probability regression loss of density map pixels is combined with the classification loss of crowd pixels.The loss function of crowd density map regression neural network is optimized by distinguishing crowd pixels from non-crowd pixels.In the end,the paper makes adequate experimental analysis and evaluation.In order to explore the effect of each innovation of SFBCC,a number of ablation experiments were carried out in the crowd density map generation algorithm,density map regression neural network design and loss function optimization.The ablation experiment results show that the methods proposed in SFBCC can improve the accuracy of the crowd counting algorithms.In addition,comprehensive contrast experiments with several crowd counting algorithms is carried out on five mainstream crowd counting data sets and SFBCC gets the lowest mean absolute error.The highest accuracy improvement of 10.6 is obtained on UCF-QNRF data set.It also achieves the lowest root mean square error on most data sets.The experimental results demonstrate that SFBCC has better accuracy and stability than other similar algorithms.