Pedestrian reloading and re-identification method for open space

Abstract

The invention belongs to the technical field of computer vision, discloses a pedestrian reloading and re-identification method for an open space, and provides a new deep neural network architecture High HRNetXt for identifying a reloaded target person. Based on the idea of dynamic optimization, a dynamic time warping algorithm is adopted to seek the distance between pedestrian skeleton key point sequences, key point detection is applied to pedestrian re-identification, and matching of the same pedestrian is completed; a ResNeXt network structure is used for improving an attitude estimation model, so that the parameter quantity is reduced by 15%, and meanwhile, the performance is effectively improved; physiological characteristics of pedestrians are extracted, so that the availability of the pedestrians in the real open world is greatly improved. Experiments show that the method provided by the invention is good in performance when facing the pedestrian reloading problem, and is superior to most of existing pedestrian re-identification models.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a pedestrian re-identification method for changing clothes in an open space. Background technique [0002] At present, with the rise of deep learning, image recognition technology has made great progress, and its recognition accuracy has even surpassed human vision. Person Re-identification (Person Re-identification) is an important application of computer vision. It breaks through the time and space limitations of a single camera, retrieves the same target person from multiple different perspectives, and realizes cross-camera tracking of the target person. Therefore, it can be used to restore the trajectory of suspects, find lost children, etc., and has great application prospects in the fields of public safety and intelligent security. [0003] The field of pedestrian re-identification has always been a research hotspot in the field of computer vision, but ...

AI Technical Summary

Problems solved by technology

[0007] (1) Existing pedestrian re-identification models rely too much on characteristic information such as clothing color when identifying pedestrians. When pedestrians change clothes, their image information will change greatly. The experimental performance under is much lower than the test results on the dataset

[0008] (2) However, the existing data set RGB-D is small in size, and the pedestrian re-identification model based on the depth map is difficult to meet the needs of high-performance depth models.

Moreover, capturing images with depth information requires special equipment, which is not suitable for large-scale deployment

[0009] (3) Existing methods require a high-precision segmenter to segment the body contours of pedestrians from the image, which is not an easy task when the color of the character's clothing is similar to the background

Moreover, this method assumes that the outline of the pedestrian's body will not change in a short period of time, only the color of the clothes will change, and the assumption is difficult to establish

[0010] (4) The existing literature uses radio signals to collect characteristic information of pedestrians, but specific equipment is required to transmit or receive radio signals, and it is unrealistic to equip large areas

[0011] (5) At present, the re-identification datasets of small and medium-sized pedestrians are difficult to be used as benchmark datasets. First, the volume is not enough to meet the needs of high-performance models; second, most datasets are virtually synthesized by computers or taken from Compared with the ideal data set PRCC from the Internet, its shooting scenes are all indoors and the background is relatively simple

[0013] Problems (1) and (5) are caused by the lack of large-scale pedestrian re-identification datasets, and it is difficult to produce large datasets. The collection of large pedestrian datasets will consume a lot of manpower and material resources, and it will involve personal Privacy issues, many countries regard citizen privacy security as part of national strategy

(2), (4) Due to its high equipment costs and related technologies, its promotion is restricted

[0016] Pedestrian re-identification is of great significance in the fields of public safety and intelligent security, but the current research is still under ideal conditions, and the implementation of the pedestrian re-identification industry is imminent. The existence of the above problems has seriously hindered the promotion of pedestrian re-identification application, especially making pedestrian re-identification impossible to break through, directly or indirectly solving the above problems will make a significant contribution to the implementation of the pedestrian re-identification industry

Images