Research On Pedestrian Counting Problem Based On Machine Learning Algorithms

Pedestrian detection technology is one of the basic technologies of intelligent traffic video analysis,and it is also a key technology in computer vision.Pedestrian detection technology is very important for pedestrian behavior analysis in pedestrian monitoring,pedestrian counting,safety monitoring,and driverless systems.The statistical process of the traditional video image pedestrian counting algorithm relies on the pedestrian target detection and tracking algorithm,the diversity of the human form in real life,the complexity of the monitoring environment,the impact of weather changes and light changes,especially the mutual obstruction between pedestrians.Overlapping and merging makes pedestrian detection have many problems to be solved.In the traditional pedestrian detection algorithm,how to design or select a feature that differs greatly from other categories and that has small differences between pedestrians and is not affected by various factors such as mutual obstruction between pedestrians and illumination conditions,and training a strong classifier is the focus of current field research.At present,the most representative and most widely used algorithm is the HOG+SVM pedestrian detection algorithm.However,the feature area information that this algorithm needs to extract is large,especially in a complex environment,it is difficult to achieve ideal detection results(when occlusion,overlap,and merge occur among people,the algorithm has a greater impact on the accuracy of pedestrian counting).With the development of machine learning algorithms,deep learning has become a focus area for scholars and researchers.Based on a large amount of training data,deep learning algorithm can automatically combine different classification tasks to learn from a large number of data and extract features.This makes it show excellent detection accuracy arobustness in target classification and target detection.However,the pedestrian detection algorithm based on the deep convolutional neural network still has some disadvantages:it is more suitable for static picture pedestrian detection,it can not associate the pedestrian's movement information,resulting in the pedestrian detection result and the counting result of each frame of the video image and adjacent The detection counts of video frames are independent of each other.This paper has conducted a thorough investigation on the pedestrian detection counting method and deep learning related technology.Based on this,it summarizes the main difficulties and problems in the practical application of pedestrian detection methods,and the traditional pedestrian detection counting method and deep learning based on Pedestrian detection methods have all focused on research and the main research results obtained are as follows.(1)For the existing traditional video pedestrian counting methods,it is necessary to extract large feature area information,and the scenes of occlusion,overlap,and merging occur among the crowds in the application scene.This paper proposes the application of pedestrian head contour features as pedestrians.Detection of Video Person Counting Methods Based on Improved GHT+KSP Algorithm.The method proposes and creates a gray-scale determiner to determine whether each circular contour detected by the GHT-type circle detection method is a true head contour.And the head contour aggregation division method is proposed to divide the multiple head contour detection results appearing in each real head region into a head true contour detection result and find the optimal contour fitting result.The human head profile is scattered in the top-down video image.This algorithm can solve the effect of pedestrian occlusion generated when using other features,and the feature extraction and aggregation method is simple.The complexity of the classification decision device is low,and it can meet the detection speed.(2)For the target pedestrian detection method of deep learning,this paper proposes a video pedestrian detection method that fuses convolutional neural network and video motion information in a sparse target scenario.The innovation of the algorithm is:combining the background modeling research method with the deep learning concentric neural network research method based on statistical learning,and integrating all the pedestrian detection result calibration boxes;presenting a pedestrian detection calibration box optimization counting method,using non-maximal The value suppression method sets the detection frame size boundary according to experimental experience,subtracts the result of the redundant result calibration frame,and determines the number of pedestrians in the frame according to the pixel area of the calibration frame.Finally,a pedestrian counting evaluation method is proposed to smooth the fused pedestrian counting results.In summary,this article has carried on the thorough research to the video person counting method,the experimental result has fully verified the validity and the practicability of the proposed algorithm.