AI has a wide range of application scenarios, and ZTE's ReID breaks the world record

Recently, ZTE Corporation has made a breakthrough in person re-identification (ReID) technology. On the three major ReID datasets Market1501, DukeMTMC-reID, and CUHK03 (these three datasets are the most influential large-scale labeled multi-target multi-camera pedestrian tracking datasets in the industry), the algorithm's key indicators mAP and Rank-1 accuracy have achieved significant improvements compared to the results previously released by various scientific research institutions and industry manufacturers, ranking first in the industry and setting a new world record.

The indicators obtained by ZTE ReID technology from three public data sets at the end of July 2019

Pedestrian re-identification refers to searching for the same pedestrian in a multi-camera network with non-overlapping viewing angles, confirming whether the pedestrian target captured by cameras in different locations at different times is the same person, that is, using computer vision algorithms to solve the recognition and retrieval of pedestrians across cameras and scenes. This technology is a major leap forward in machine learning in image perception after face recognition technology. It can recognize pedestrians based on their clothing, body shape, hairstyle, behavior and other information, shortening the time for traditional manual video review and locking the trajectory of key personnel from dozens of days to minutes and seconds.

In the field of pedestrian re-identification technology research, the first hit rate (Rank-1 Accuracy) and the mean average precision (mAP) are the core indicators for measuring the level of the algorithm. ZTE innovatively proposed a multi-module and multi-granularity joint feature extraction network based on this technology, which effectively solved the complex situations such as side faces, occlusions and missing faces caused by camera shooting angles, color differences, and light intensity, and significantly improved the network feature matching performance. It achieved a significant improvement in the two key indicators of Rank-1 (first hit rate) and mAP (mean average precision) on the three major data sets, ranking first and setting a new world record. (Note: This result was achieved without using any time series information and without re-ranking optimization)

ZTE ReID technology partial test results

ZTE has always insisted on increasing investment in research and development. In addition to 5G and other communication businesses, it has also been deeply involved in the video and computer vision fields, focusing on intelligent video business. It has a deep accumulation of technical algorithms such as target detection, multi-target tracking and various target recognition, and has made continuous innovations and breakthroughs. As early as 2017, it ranked first in the multi-target tracking (MOT Challenge) algorithm evaluation. In practical applications, pedestrian re-identification technology can be combined with face recognition technology to make full use of existing equipment and infrastructure to achieve continuous tracking and trajectory restoration of retrieval objects across cameras, enhance the spatiotemporal continuity of data, and raise the cognitive level of artificial intelligence to a new level.

As pedestrian re-identification technology becomes increasingly mature, it will promote the continuous development of intelligent perception applications of personnel on a large scale, bring great application value and practical significance to the fields of commerce, transportation, finance, public safety, etc., and allow artificial intelligence to serve more scenarios and more users.