Research On Person Re-identification Algorithm In Intelligence Visual Surveillance

The goal of person re-identification(Re-ID)is to determine whether the pedestrian of interest is the same person as the pedestrians captured at different times and places across the camera.In recent years,person re-identification technology has made good progress,but there are still many challenges in designing an effective re-ranking algorithm and high-precision cross-domain person re-identification algorithm.On the one hand,most of the current Re-ID algorithms use the appearance characteristics of pedestrians to distinguish,but the initial ranking list obtained by the Re-ID algorithm contains many wrong matches because of the viewpoint,the posture,the lighting and partial occlusion.And many algorithms use Euclidean distance for feature matching,which has certain limitations.On the other hand,the single-domain Re-ID algorithm has achieved very good accuracy,but the performance of cross-domain person reidentification is poor.Obviously,it is unrealistic to establish a database for each scene.Cross-domain person re-identification is closer to the practical application.The target domain label is not available,which makes the current Re-ID algorithm cannot be directly applied to cross-domain person re-identification.Faced with many of the above challenges,this paper attempts to adopt a re-ranking algorithm based on extended kreciprocal nearest neighbors,a cross-domain person re-identification algorithm based on similarity preserving generation adversarial network and multi-granularity network to solve some problems existing in the current person re-identification algorithm.The main research work is as follows:1)The ranking list similarity distance can be effectively calculated by single matrix multiplication and addition.The ranking list similarity distance is more accurate than that of the traditional distance metric method.The probe set and gallery set are combined into a single matrix.And the ranking list similarity distance can be effectively calculated from the original distance list through single matrix multiplication and addition operations.2)A re-ranking algorithm based on extended k-reciprocal nearest neighbors is proposed.Taking full advantage of the constraints of k-reciprocal neighbors,a twolevel extended M k-reciprocal neighbor distance calculation method is designed.Given a probe image,replace the probe with its two-level extended reciprocal nearest neighbor to ensure that all the extended neighbor sets are correctly matched.The final distance is computed by the mean of the corresponding neighbor set.Experiments show that the proposed method further improves the accuracy of Re-ID algorithm.3)A framework for cross-domain person re-identification is proposed.The similarity preserving generation adversarial network is used to transform the sourcedomain training image into the target-domain style image and learn the target-domain style characteristics;the source-domain image and the converted target-domain image are sent together into a multi-granularity feature extraction network.The experiment shows that the method has good performance in both single-domain and cross-domain person re-identification,and further improves the accuracy of cross-domain person reidentification.