A Crowd Counting Method Based On CNN And Trajectory Prediction

In recent years,with the continuous development of computer technology,intelligent video surveillance system has been in shopping malls,schools,railway stations and other public places for a large number of applications to ensure the orderly and stable operation of the community.Crowd counting is a practical research direction in the field of intelligent video surveillance.It is also one of the hotspots and difficulties in the field of computer vision.Accurate crowd counting is of great significance in the public security prevention and control,business information collection and the allocation of social resources and facilities.Crowd counting is difficult to get accurate statistics due to shading,shadows and changes in crowd density.To solve the above problems,this theses analyzes the applicable range of common pedestrian feature extraction algorithm,and divides the uneven distribution of crowd according to the density.It proposes a crowd counting algorithm which combines pedestrian recognition with feature regression legitimately.For sparse crowd,the convolutional neural network is used to recognize pedestrian targets.While,feature regression algorithm is used in dense crowd so that the algorithm can cope with the complex crowd scene.Selective search algorithm is used in HSV color space to locate the people to avoid the interference of illumination variation,rain and fog.Then convolutional neural networks features are extracted with grid loss function to avoid the occlusion issue.Crowd density map features are extracted to train the support vector regression models to estimate the number.In the video,the trajectory is predicted by the Markov model based on the trajectory of the blocked pedestrian.When the pedestrian is no longer obscured and appears at the predicted trajectory position,the pedestrian target is locked and counted as the number of the obscured frame.Experiments are conducted on datasets PETS2009 ? UCSD in Caffe framework.The experimental results show that the proposed algorithm improves the accuracy to some extent in comparison with other algorithms.