Research On Human Detection Technology Based On Multi-feature Fusion

With the development of digital image processing related technologies,image processing capabilities have been significantly enhanced,which has accumulated more technical foundation for human detection under static images.Human body detection is to obtain surrounding scene information through computer vision equipment,and then to detect the target human body in the image or video through the human body detection algorithm,and to make accurate positioning.The technology is widely used in intelligent video surveillance,smart mobile device photography,car-assisted driving,smart city intelligent transportation and other fields.However,due to the interference of non-uniform illumination and partial occlusion in the actual application scenario,the recognition effect under the interference condition is not very satisfactory.In order to propose a set of human detection methods that are more robust to non-uniform illumination and partial occlusion interference in real-world scenarios.This paper firstly extracts and fuses multiple features of human images.The Locality Sensitive Histograms of Oriented Gradients(LSHOG)extracts the global features in the image into each point.The local sensitive gradient direction histogram is combined with the Gray Level Co-occurrence Matrix(GLCM),which also has the ability to describe global texture features,to further enhance the global representation ability of the feature.Secondly,the feature is reduced by the under-complete Sparse Auto-encoder(USAE).In order to prevent the gradient dispersion and over-fitting of the auto-encoder(AE),a loss factor is added to the model loss function to suppress some nodes.Finally,in this paper,a human detection network model based on convolution and sparse self-encoder is proposed.This model uses the Convolution Auto-encoder(CAE)to extract the high-level features of the original image and combine it with the underlying features of the USAE dimensionality reduction.The Extreme Learning Machine(ELM)is used to classify and identify the combined features.In this paper,15000 samples cut from INRIA data set and Daimler data set are tested.The results show that the proposed human body detection model is effective.