Research On Convolutional Neural Network Based Pedestrian Gender Recognition Method

With the rapid development of internet,cloud computing,and multimedia technologies,digital video surveillance systems have been widely deployed in various areas.In these digital video surveillance systems,pedestrian identification plays a very important role for public safety.Hence,the pedestrian intelligent surveillance systems have been emerging.Pedestrian gender recognition,as an assistant method for pedestrian identification technology,has important research value for the long-distance target identification.However,pedestrian gender recognition is a very challenging problem.On one hand,the existing surveillance systems are hard to capture clear face images in the long-distance conditions,thus,it is difficult to recognize the gender of pedestrian based on the face information.On the other hand,the viewpoint variation,background clutter,blurred images,illumination change and object occlusion are frequently encountered in the pedestrian images,seriously affecting the pedestrian gender recognition.Consequently,based on the characteristics of the pedestrian images and the convolutional neural networks(CNNs),two effective pedestrian gender recognition methods are investigated in this thesis.The major contributions of this thesis can be described as follows:1.The HOG feature is able to effectively capture the local feature information of a person,while the deep learning feature is a more advanced semantic feature,having better robustness to the view change,illumination change,object occlusion and background clutter.To effectively explore the advantages of both HOG feature and deep-learned feature,an effective HOG-assisted deep feature learning approach is proposed for the pedestrian gender recognition.Firstly,the proposed HOG-assisted deep feature learning approach simultaneously performs the deep-learned feature and weighted HOG feature extraction on the input pedestrian image,then,the obtained features are further fused together to obtain a more effective feature.Extensive experiments have been conducted on multiple pedestrian datasets,showing that the proposed HOG-assisted deep feature learning is able to effectively represent the pedestrian gender attributes,and consistently outperforms multiple state-of-the-art pedestrian gender recognition methods.2.Different image representations may produce different gender prediction scores through a CNN model.In order to obtain a superior prediction scores,an effective decision information fusion learning method is proposed for pedestrian gender recognition.Firstly,the Prewitt gradient operator filter is exploited to transform the color pedestrian image into corresponding gradient magnitude image.Secondly,two parallel CNN feature learning modules are constructed to learn the feature vectors from the input original image and its gradient magnitude image.Finally,a decision information fusion module is designed to perform weighted summarization to the prediction scores computed from corresponding feature vectors.Comprehensive experiments on multiple pedestrian image datasets have illustrated that the proposed decision information fusion learning method is able to obtain a more accurate prediction scores,and is superior to multiple state-of-the-art pedestrian gender recognition methods.In summary,to some extent,the proposed methods in this thesis broaden the thinking for pedestrian gender recognition and provide the technical support for the application of the pedestrian intelligent systems.