Deep Learning Based Surveillance Pedestrian Detection

Pedestrian detection is an important research area in the field of computer vision and pattern recognition.It has extensive application in many areas,such as intelligent transportation,intelligent video surveillance,self-piloting automobile and intelligent robots.For example,intelligent video surveillance focus on movement and behavior of pedestrians,to achieve tracking of pedestrian and analysis of pedestrians' behavior,the first thing to do is detect the pedestrians in videos.Traditional pedestrian detection solutions are mainly based on statistical classification methods.By manually designing features with small intra-class differences and large differences among classes and training classifiers,the detection problems are transformed into the dichotomous problems of human and non-human.At present,commonly used features include Gradient Histogram(HOG)and Scale Invariant Feature Transform(SIFT)feature.However these hand-designed features do not have strong characterization abilities for pedestrian diversity and have some limitations in terms of robustness,furemore designing these features requires a solid,professional foundation.In recent years,Deep Learning(DL)has been widely used in computer vision problems and has achieved great success.DL uses Convolutional Neural Network(CNN)to automatically learn the feature of the target,and this feature has strong robustness.In this paper,we carry out our research on pedestrian detection based on deep learning algorithm,aimed at improving the detection accuracy of small-scale pedestrians.The main innovations of this paper are as follows:(1)Based on the structure of CNN,the context information is modeled to realize the pedestrian detection model with scale-adaptive context information.First of all,a multi-scale pedestrian detector is trained using templates of different sizes,the multi-scale detector obtains the classification and the bounding box regression information by predicting the heatmap.According to the characteristics of CNN receptive fields,it is possible to add different sizes of contextual information to different scales of pedestrians.Specifically,small-scale targets add huge contextual information,whereas large-scale targets only need a small amount of contextual information.(2)CNN features are fused to construct the feature pyramid structure to enhance the use of different layers of CNN features and improve the detection of small-scale pedestrians.CNN's high-level features have large receptive field and rich semantic information,which is suitable for detecting large-scale pedestrians.while the features of lower-level features are not so robust and its characterization ability is not strong.However,abundant detailed information is retained to facilitate the detection of small-scale pedestrians.In this paper,we will make full use of the features of different layers of CNN.Firstly,we set up detectors on different features of CNN and detect the pedestrians with different scales with different size of templates.Secondly,we strengthen the characterization of low-level features through the fusion of high-level and lowlevel features to improve the accuracy of pedestrian detection.