Pedestrian Detection Based On Enhanced Depth Feature Representation

With the development of big data technology,artificial intelligence has ushered in unprecedented development opportunities.The degree of integration of artificial intelligence with society and human life has also continued to deepen.The emergence and application of artificial intelligence technology not only greatly enriches people's leisure and entertainment life,but also promotes the construction and adjustment of new industrial system in transportation,traditional manufacturing,medical,education and other industries.At present,the types of data processed by artificial intelligence technology mainly include text,voice,image and video.According to the research of Cisco in the United States,it is estimated that the global online video traffic will account for about 82% of the total traffic by 2022,and the online data will be increasingly dominated by video data.Computer vision,which mainly deals with image and video,has more abundant application scenarios and commercial value than other artificial intelligence technologies,which mainly deal with text or voice.According to the data from Tsinghua University,at the technical level,the application of domestic computer vision in China accounts for 34.9% of the entire application field of artificial intelligence,which has become an important support for the development of various industries,including automatic driving,augmented reality,security,medical image analysis,finance,unmanned retail,etc.Human based detection is one of the core contents of computer vision Pedestrian detection is that the computer determines whether there is a pedestrian based on a given image or video,and gives the specific location information of the pedestrian.Pedestrian detection is the basis and premise of pedestrian tracking,behavior analysis,gait analysis,pedestrian identification and other research.It provides important technical support for video surveillance,vehicle assisted driving,intelligent robot and other fields.Human not only has the general object attribute but also has the particularity of multiple changes within the class.Human's special attribute is also the difficulty of pedestrian detection.The detection method based on traditional artificial design feature detection method,from the initial use of simple features to artificial design feature descriptors,the ability of feature representation is gradually enhanced.In recent years,the deep convolution neural network with strong learning ability has made great progress in the field of computer vision,but there are many problems in the field of pedestrian detection,such as the variety of pedestrian types,the complexity of background,the wide range of pedestrian scale changes,etc.Therefore,this paper increases from the basic components of convolution neural network,the guidance between feature map channels,the fusion of different fine-grained features,etc.The feature representation ability of strong network improves the performance of pedestrian detection.The following work has been completed in this paper:In this paper,an adaptive pooling method is proposed to replace the single pooling method commonly used in convolutional neural network,which can increase the information content of the feature map.The conventional pooling method is single pooling,which is not suitable for pedestrian non rigid characteristics,especially the multi-layer pooling operation will cause serious loss of information in the deep feature map.It causes the problem of missing detection of small-scale pedestrians and the problem of reducing the positioning accuracy of large-scale pedestrians.The proposed adaptive pooling method can retain the edge features of the object contour(gradient change obvious area),and improve the information content of the feature map under the same resolution.The proposed adaptive pooling method combines multiple single pooling methods through learning mechanism to increase the amount of information.Based on the learning mechanism,the connection method can adjust the contribution of different regions in the feature map,and realize different regions to use the corresponding pooling method.Reduce the loss of useful feature information.It has been tested on several pedestrian detection data,and the detection performance has improved significantly.Secondly,a feature adjustment method based on channel contribution is proposed.At present,most of the existing deep neural networks mainly enhance the spatial coding ability of convolution operation to improve the ability of feature representation.Due to the different contribution of each channel to pedestrian recognition,the channel features with the greater contribution are enhanced,and the channel features with the smaller contribution are suppressed to adjust the features and improve the representation ability of the features.In this paper,a reduction adjustment(RA)module is proposed to establish the correlation between the channels.The contribution degree of the channel is represented by the weight value,and the contribution degree is used to adjust each characteristic value in the corresponding characteristic map of the channel.It enhances the useful feature information and suppresses the irrelevant information.In experiments on multiple datasets,RA module can improve the accuracy of pedestrian detection by adjusting the network feature representation.Finally,a multi feature fusion method based on global features is proposed.The scale distribution of pedestrians is widely distributed,but the number of pedestrians corresponding to each scale is is unevenly distributed,and the number of small-scale pedestrians is far greater than that of large-scale pedestrians.At present,in the detection framework based on convolutional neural network,high-level semantic information can be obtained in the feature map of depth network,which is conducive to large-scale pedestrian detection,but not conducive to small-scale pedestrian detection.Because the information content of different fine-grained feature graphs is different,the information content is complementary by fusing multi-layer feature graphs.The innovation proposed by introducing global context information can significantly improve the information of small-scale pedestrians.The Global feature Faster R-CNN pedestrian detection framework is proposed and tested in multiple data sets.The results show that the proposed method has significantly improved the detection of multi-scale pedestrians,and the detection framework is accurate in positioning,which can tightly surround pedestrians.Compared with various detection methods,the method proposed in this paper has significantly improved both the detection effect and the positioning accuracy.The proposed method is flexible and can be used not only for pedestrian detection but also for general object detection,and has good practicability and generalization.