Person Re-identification Based On Feature Fusion And Metric Learning

Monitoring systems play a vital role in public safety.With the increasing scale of monitoring systems,traditional manual monitoring methods have been unable to adapt to social development.Therefore,intelligent video surveillance has attracted more and more attention from researchers in recent years.In the intelligent monitoring system,the person re-identification part plays an irreplaceable role and is the link of the non-overlapping camera target trajectory.Although the person re-identification algorithm achieves certain accuracy on some specific data sets,illumination,occlusion,pedestrian attitude change,and perspective change are still challenges in the field of pedestrian recognition.Based on the analysis of existing pedestrian recognition algorithms,this paper studies both traditional algorithms and deep learning algorithms.In the traditional algorithm,a person re-identification algorithm based on feature fusion and subspace learning is proposed.before the feature extraction,the image is first preprocessed to enhance the image processability.After the pre-processing,local features and overall features are extracted for the pedestrian image.The overall feature combines HOG features and HSV color histogram features as an overall feature.The local features take the extraction of CN color features and two scale SILTP features within the sliding window.In order to obtain multi-scale features,the original image is separately downsampled twice,and then the above features are extracted separately.After the feature extraction,the extracted feature is transformed into the nonlinear space by the kernel function,and then a subspace is learned in the nonlinear space.Finally,a measure matrix is learned in the sub-feature space to measure the similarity.In terms of deep learning,a multi-loss fusion person re-identification algorithm for multi-granular component alignment is proposed,the backbone network with feature extraction is obtained by fine-tuning the pre-training network,and two branches are connected after the backbone network,and the pedestrian characteristics extracted by the backbone network are processed differently.In order to deal with the problem of pedestrian image misalignment in the pedestrian detection frame,the alignment of the local feature with the minimum total distance is found by dividing the image level into different parts and then dynamically matching the local parts from top to bottom.The multi-granularity combination of pedestrian images is also an effective means to improve the recognition rate.The pedestrian image is divided into two parts,and the loss of each part is calculated separately.Considering the particularity of the pedestrian re-recognition algorithm,this paper will combine the three-element loss and the classification loss multi-loss fusion method,and calculate the sum of the losses of different depth features as the overall loss.when testing,a simple distance function was used to calculate feature similarity and reordering calculation similarity.The proposed traditional algorithm has certain advantages in the test on three data sets.The deep learning algorithm also shows certain advantages in the recognition rate of two large data sets.