Research On Multi-modality Pedestrian Re-identification Algorithm And Feature Retrieval

With the continuous development of cities and the continuous improvement of the level of urbanization,the issue of urban safety has been more and more concern.Helped by the complete video surveillance system of "Skynet Project",police officers can trace suspects and missing people through video data and restore their location.However,if such retrieval relies solely on labor,we have to spend a lot of energy on finding and distinguishing the identity of pedestrians,which greatly reduces the efficiency.The pedestrian re-identification(Re-ID)task is to search and return several results most likely to be the given query person in the video database,make quick retrieval and match the identity of the pedestrian.However,the task faces difficulties including pedestrian appearance and posture changes in multiple shooting scenes,objective gaps in different modalities,and possible dimensional disasters in feature retrieval.This article summarizes several research hotspots of pedestrian re-identification at the current stage,and starts from two perspectives of multi-modality pedestrian feature extraction and feature retrieval acceleration methods as follows:(1)Research on feature extraction network for single-modality Re-IDFor the Re-ID task in single modality,this paper proposes a multi-level slice-based network(MSN),with basic structure of two-branch structure,which are global feature extraction branch and local feature extraction branch.Both branches utilize multi-level information in different levels and combine different sliced regions.Experiments show that the proposed MSN achieves excellent results on several mainstream single-modality Re-ID datasets,proving the effectiveness of MSN and joint objective function.(2)Research on feature extraction network for cross-modality Re-IDFor cross-modality Re-ID task,this paper proposes MSN based on channel Zero-padding.After performing channel Zero-padding on the input,the cross-modality network can be characterized by single-stream input structure as MSN.Also the design of the joint loss function enables the network to overcome the modality gap.Experiments show that the modified MSN achieves excellent results on cross-modality Re-ID dataset,proving the effectiveness of network and joint objective function.(3)Accelerating Re-ID feature retrieval methodFor feature retrieval acceleration,this paper analyzes two types of methods using K-D tree and product quantized inverted index.The method based on K-D tree makes accurate retrieval,but the feature dimension will affect the retrieval efficiency,while the size of tree structure has not decreased and even increased.Due to the existence of quantization errors,the method based on the product quantized inverted index will decrease in accuracy but greatly shorten the retrieval time,and the space occupied by the quantized coding features will be greatly reduced.Experiments show that both methods above can accelerate Re-ID feature retrieval compared with force retrieval.