Crowd Counting Aiming At Large-scale And Uneven Distributed Pedestrians

With the rapid growth of China's population and the development of economy,more and more people are pouring into the first-and second-tier cities to work and live.As a result,congestion,traffic jams and other incidents caused by the diversification of crowd and large-scale gathering of people become a common scene.Although the government has deployed a lot of security guards in managing the crowd and installed more and more surveillance cameras to monitor and manage the quantity and flow of people,they have a limited influence.Traditional monitoring is facing problems like: not accurate enough,not efficient enough and unexpectedly expensive.Therefore,monitoring and counting pedestrians with the help of artificial intelligence has become a hot topic.This paper proposes a crowd counting algorithm based on single image after a thorough research.Our algorithm end-to-end maps the captured image to its crowd density map through a deep convolutional neural network and calculates the number of pedestrians.Facing the two challenges in crowd counting: large-scale pedestrians are difficult to identify and the population is unevenly distributed.Our algorithm proposed two solutions: for the former problem,our algorithm designed a hybrid dilated convolution module to increase the model's perception field and improve the model's ability to recognize large-scale pedestrians.For the latter,our algorithm designs an attention module with encoding-decoding structure.By gradually learning semantic features in the image through multiple feature encoding units and gradually recovers the feature map through multiple feature extraction units,our algorithm merged the details of previous network and obtain an attention area.By doing so,the irrelevant areas are excluded and a more accurate crowd counting result can be finally achieved.In addition,different from other algorithms,our algorithm generates a high-precision crowd density map with the same resolution as the input image by gradual expansion of the feature.This paper first verifies the validity of hybrid dilated convolution module and the attention module on dataset Shanghai Tech Part_A.Compared with conventional CNN,network with hybrid dilated module has a mean absolute error of 124.6,which is 17.9% lower;network with hybrid dilated module and attention module has a mean absolute error of 79.5,which is 47.6% lower.What's more,the crowd density map generated by CAA Net proposed by this paper has a SSIM of 0.83 and has an absolute error of 19 on large scale pedestrian(less than 200 people in one picture).This paper also does experiments and tests on Shanghai Tech Part_B and World Expo 10 datasets with sparser crowd.It is indicated that CAA Net's mean absolute error and root mean square error are 22.1,23.5 and 3.3,4.3,respectively.It is proved that our proposed algorithm can obtain a high counting accuracy and robustness and is capable of generating high quality crowd density maps.