Research On Pedestrian Detection And Re-identification Method Based On Fusion Attribute Feature

In the intelligent video surveillance system,Person re-identification is a key step,which can help the system quickly and accurately locate the position of the same person under different lenses.Person re-identification under complex background has become one of the research hotspots.Person re-identification methods are to match the person image to be tested with the person image that has been clipped in the matching library.However,the person in the actual video surveillance are not marked,and the person images have occlusions.Different lighting intensity and shooting Angle will also affect the image quality.Therefore,accurate target detection should be carried out before person re-identification.This thesis combines person target detection and person re-identification on the basis of deep learning method,and completes the following three main research works:(1)A person target detection method based on improved SSD(Single Shot MultiBox Detector)algorithm is proposed.According to the excellent performance of convolutional neural network in the field of object detection,some classical network models are compared and analyzed.Finally,SSD model is selected as the basic network model,and on this basis,the backbone feature extraction network of the model is improved.Secondly,in order to improve the detection effect of small and medium-sized targets,the information of feature layers at different scales is extracted to carry out multi-scale feature information fusion,effectively suppressing the background noise introduced by high-level features.The test results show that this method can guarantee the detection speed and improve the detection accuracy significantly.(2)An end-to-end person re-identification model is proposed which combines person target detection and person re-identification tasks.Based on the improved SSD algorithm proposed above and further improved on this basis,two joint networks are established for two person scenes under two target scenes of person re-identification,and joint training is carried out to obtain the target detection and re-identification system.Finally,the target detection and re-identification system is used to identify the target,and the result of target detection and reidentification is obtained.Experiments on Market1501 and Duke MTMC-re ID data sets show that the algorithm is effective.(3)A new attribute feature fusion method is proposed to improve the performance of person re-identification model.Multi-task learning is carried out on the improved SSD network to better capture local features,so as to realize the effective combination of attribute recognition and person re-identification.The network can generate adaptive weights corresponding to each attribute,and combine all attributes with global features through weighted summation to complete the task of person re-identification.This method not only pays attention to the overall appearance of person but also pays attention to the detailed regional information of person to give a more comprehensive description of person.Experimental results on Duke MTMC-re ID and Market-1501 data sets show that the proposed method is effective.