Pedestrian Detection Method In Traffic Monitoring System

In pedestrian surveillance video,the traditional pedestrian detection method is manual recognition.With the increase of pedestrian video data,this traditional method not only consumes a lot of manpower,but also has very low efficiency and is prone to missed detection.If the machine is used to complete the process,the detection efficiency will be extremely improved and the consumption of human resources and financial resources can be reduced.Pedestrian detection technology can solve these problems in traditional pedestrian monitoring by giving the machine the ability to intelligently identify pedestrians and track pedestrians.Therefore,pedestrian detection has rich commercial application value in many fields,such as advanced assisted driving,intelligent security monitoring,and intelligent robotics.This thesis mainly studies the pedestrian detection method in traffic monitoring system.The concrete contents are as follows:In view of the HOG pedestrian detection method,the pedestrian detection accuracy is not high under the illumination variation and background complexity.Based on the HOG features,this thesis introduces the HOP features to construct the HOPG features,which can identify and locate more structural information,and is more robust to illumination changes and complex backgrounds.The input image is first divided into local regions,and the gradient and phase congruency magnitude and direction are calculated relative to its neighborhood in each pixel of the input image.The histogram of oriented gradient and the histogram of the oriented phase are computed in each local region and combined to each other.These histograms are concatenated to form the HOPG features that representing the input image.These features are finally fed to the trained SVM classifier to determine if it is a pedestrian.Experiments show that the pedestrian detection method based on HOPG features has better detection performance than the pedestrian detection method based on HOG features.Aiming at the problem that the ACF pedestrian detection method generates more false detection windows,this thesis introduces the convolutional neural network VGG-16 based on the ACF method,and trains a pedestrian target classification model using the fine-tuned VGG-16 network.The model is capable of extracting high-level abstract features that benefit pedestrian detection.Firstly,a rough AdaBoost classifier model is trained by ACF extracted features to classify test samples to obtain candidate detection windows.Then,a fine AdaBoost classifier model is used to classify the candidate detection window by using the features extracted by the fine-tuned VGG-16 network.Finally,the NMS algorithm is used to eliminate the redundancy pedestrian detection box.Experiments show that the pedestrian detection method based on the combination of ACF and VGG-16 reduces the rate of missed detection and improves the detection rate compared with the pedestrian detection method based on ACF.The pedestrian detection method based on ACF and convolutional neural network is generally divided into two steps.The classifier of ACF feature training is used to generate candidate regions,and the classifier for convolutional neural network training is used to classify candidate regions,although the detection performance has further improvement,but there is still a need for improvement.For example,the convolutional neural network trains the classifier through the extracted pooled layer features,ignoring the regional score and the features of the CNN inner layer.To overcome this shortcoming,multi-layer channel features are constructed for pedestrian detection.It first integrates HOG+LUV with each layer of CNN into a multi-layer image channels.Based on the multi-layer image channels,a multistage cascade AdaBoost is then learned.The weak classifiers in each stage of the multi-stage cascade are learned from the image channels of corresponding layer.The non-pedestrian detection window is excluded using the weak classifiers of each stage.Experiments show that the detection performance of the pedestrian detection method based on the multi-layer feature is further improved compared with the ACF+VGG-16 and HOPG methods.