Research And Implementation Of A CNN-based Crowd Counting Method

With the Rapid growth of global population and human social activities in recent years,crowded scenes appear frequently in public places all over the world,like traffic centers,cultural and sports centers and religious places.This brings security risks to public safety.In order to prevent dangerous situation like stampede,such public places most have security surveillance cameras.But human monitored surveillance are low-efficient and high-cost.Therefore,automatic,efficient and precise dense crowd counting with the help of computer vision and artificial intelligence technology has become a major research field for researchers.Crowd counting is also useful in fields like public space design,transportation facility planning and intelligent city.Crowd counting is a challenging problem in the field of computer vision.Its purpose is to take in a crowded image as input,and output the corresponding density map and the number of people in the image,which show the dense level and distribution of the current scene.The challenge of the problem includes: blocking of people in dense crowds,which increases the difficulty to identify;people in far distance contain insufficient pixels to detect;the differences in camera perspective and distance,which cause object scales vary significantly within one image or between images,and difficult to adapt to different scenes.In order to address these problems,a crowd counting algorithm exMCNN is proposed,which implements the mapping from dense crowd images to corresponding density maps and estimated number of people.The main work of the paper includes:1.Three convolutional neural networks with different convolutional filter scales are implemented to solve the problem of perspective distortion and scale inconsistency caused by differences of camera perspective and distance.These CNN networks are used as basic regressors to calculate estimated density maps from the input image respectively.The regressors have the same architecture,each is suitable for a particular scale of crowd object.2.A feature extractor based on VGG architecture is implemented to address the problem of scale inconsistency between images and within image.It extracts features of every places of the images.The final part of the algorithm combines the output of the regressors and the feature map from the feature extractor,and calculate the final densitymap.The density map is able to show the distribution of the crowd in the image,and the total number of people is the sum of every pixel of the image.3.The proposed algorithm is tested on several major crowd counting datasets,including ShanghaiTech,UCF_CC₅0 and Mall.The experiments compared the algorithm with formal crowd counting methods,and proved that the algorithm has state-of-the-art performance.Also,feature map visualization during and after training is shown to prove the effectiveness of the network design.Further more,the problem of scene transfer is introduced and the possible solutions are discussed.Finally,a real-time crowd counting program based on exMCNN method is introduced.