CN-115797398-B - Pedestrian track flow multistage clustering algorithm considering multi-camera information

Abstract

The invention discloses a pedestrian track flow multistage clustering algorithm taking multi-camera information into consideration, and relates to a semi-supervised video target geographic flow multistage clustering algorithm. The method adopts geographical stream layering representation to a video target track based on a stream space to obtain geographical stream sub-segment sets of different levels, adopts a camera number as a label to respectively perform semi-supervised clustering on the stream sub-segment sets of different levels to obtain a cluster center set of corresponding levels, and effectively describes the general trend of a large number of geographical streams by using the cluster center. The method overcomes the defect that the traditional algorithm can only perform single-level clustering and does not consider camera information. A series of experiments demonstrated the effectiveness of the algorithm.

Inventors

• XIE YUJIA
• WANG WAI

Assignees

• 南京财经大学

Dates

Publication Date

20260505

Application Date

20220922

Claims (3)

1. 1. The pedestrian track flow multi-level clustering method considering the multi-camera information is characterized by comprising the following steps of: (1) The method comprises the steps of inputting a group of cross-camera video target tracks, and carrying out multi-level stream sub-segment decomposition according to the number of the video targets passing through cameras; (2) Semi-supervised clustering is carried out on the sub-segment sets of different levels respectively, thereby obtaining The group clustering center set comprises that for a flow sub-segment set under the same level, q samples are arbitrarily selected as initial clustering centers, distances between other samples in the flow sub-segment set and all the clustering centers are calculated, and each sample is attributed to the clustering center with the nearest distance: (12) (13) Where s represents an arbitrary sample, c i represents the ith cluster center Finally, checking whether the marks of each sample are the same as the marks of the current clustering center, and if so, performing sample approaching operation: (14) Otherwise, sample distance operation is carried out: (15) the samples here are marked as camera numbers through which the stream subsections pass; The said process The cluster centers of the group are as follows: . Wherein, the (9) (10) (11) Wherein, the For the track stream to span the number of cameras, Representing the track flow as the number of crossing cameras A set of cluster centers at the time of the time, Represent the first And clustering centers.
2. 2. The multi-camera information-based pedestrian track stream multi-level clustering method as claimed in claim 1, wherein the video object of step (1) is represented in a stream space: A video track OD stream is composed of a series of stream subsections, video targets Expressed in the stream space as: (1) Wherein, the (2) Wherein, the And Representing the coordinates of the map and, Representing a time stamp; To obtain And Taking the coordinate of the entering area and the coordinate of the exiting area of the video target in each camera as the storage moment of the streaming subsections, Constructing a mapping model by adopting a homography matrix method, and assuming that the image coordinates of the exit area/the entry area are Geospatial coordinates of Then And (3) with The homogeneous coordinates of (c) can be expressed as: (3) (4) set the mapping matrix as Then And (3) with The relation of (2) is: (5) The mapping matrix can be transformed from the image plane to the geospatial plane by scaling, translation and rotation The method is divided into: (6) In the formula, Is a scaling factor; translating the transformation matrix for the camera; for a3 x 4 dimensional rotation transformation matrix: (7) (8) Wherein, the 、 The product of the physical focal length of the lens and the dimension of the sensor in the horizontal axis and the vertical axis directions of each unit is respectively expressed; 、 offset of the image imaging center relative to the main optical axis on the horizontal axis and the vertical axis respectively; 、 、 respectively representing the coordinate system in the physical space A shaft(s), A shaft(s), A rotational relationship in the axial direction; Representing a translational relationship between coordinate systems; When using the homography, the camera view plane in the geospatial is assumed to be horizontal, i.e. at this plane The mapping of image space to geographic space can thus be regarded as one plane to another, to simplify the calculation In (a) and (b) And (3) with Middle-indicated winding With rotation of the shaft Removing the homography matrix The simplification is as follows: (9) according to the matrix The geospatial coordinates of the exit/entry area of the video object can be found.
3. 3. The multi-stage clustering method of pedestrian track streams considering multi-camera information according to claim 1, wherein the step (1) is characterized in that, for the situation that the same video target track exists in multiple camera views, for performing hierarchical clustering on stream subsections of different levels, the track stream is represented in a multi-layer manner: Is provided with Through the process of , , The number of layers is 5; Passing through , The number of layers is 3, Passing through The number of layers is 1; At the same time contain the slave Start to Ending a sub-segment of the stream with a layer number of 3; At the same time contain The lower stream sub-section has a layer number of 1.

Description

Pedestrian track flow multistage clustering algorithm considering multi-camera information Technical Field The invention relates to the field of track clustering under the conditions of video-geographic scene data fusion organization, track OD flow space and flow mode, in particular to a pedestrian track flow multistage clustering algorithm considering multi-camera information. Background Along with the development of a monitoring video system from a single camera to a multi-camera network, the traditional video target track clustering algorithm is poor in clustering effect due to the fact that camera position information and a view range are not considered. The video target positioning device and the position tracking technology are advanced, a large amount of track data is generated, and the movements of people, vehicles and animals in various applications are recorded, so that the video target positioning device and the position tracking method have the characteristics of convenience in deployment, visual information and rich media expression form. Track clustering defines similarity measurement of tracks in space-time dimension, so that each track is distributed to reasonable clusters, and how to effectively cluster video target tracks is always a research hot spot. The existing video target track clustering method has certain defects in the analysis object layer and the analysis method layer. The method is characterized in that the method is used for clustering a small-scale data set under a single camera, and is poor in real-time performance and accuracy in a large-scale cross-camera track data stream. The path of the actual track of the video target is complex, and high calculation cost is needed in cluster analysis. Furthermore, dimensions between different target trajectories are difficult to unify. The flow space tends to ignore the actual path or short-term single-target motion state, and the flow or interaction i of the target object between two starting and ending points with geographic positions is mainly analyzed, so that the flow space is not limited by geographic space distribution and hierarchical structure, and the interaction relation characteristics of different regional sites are reflected. The stream space and the video target track are combined and analyzed, so that a video target analysis means can be improved, the analysis value of the existing video target can be improved, and the analysis capability of the multi-camera video target in the geographic space can be improved. The research is based on a geographic flow space, and a video moving target is used as a research object to research a multi-level pedestrian track flow clustering algorithm. In summary, the invention provides a semi-supervised trajectory stream clustering algorithm taking camera space information into consideration, which uses the serial numbers of cameras as labels for the trajectory stream subsections in the camera view, and uses the serial numbers of a camera above and a camera below the blind area trajectory stream subsections of the cameras as labels to perform hierarchical clustering on the stream subsection combinations of different levels. The clustering effect and the time efficiency of the algorithm are shown through experiments, and the effectiveness of the algorithm is proved. Disclosure of Invention Aiming at the defects in the prior art, the invention aims to provide a pedestrian track flow multistage clustering algorithm taking multi-camera information into consideration, which adopts geographic flow layering representation on a video target track based on a flow space to obtain geographic flow sub-segment sets of different levels, and adopts camera numbers as labels to respectively carry out semi-supervised clustering on the flow sub-segment sets of different levels to obtain a clustering center set of corresponding levels, so that the overall trend of a large number of geographic flows is effectively described by the clustering center. The method overcomes the defect that the traditional algorithm can only realize single-level clustering and does not consider camera information, and a series of experiments prove the effectiveness of the algorithm. In order to achieve the aim, the invention is realized by the following technical scheme that the pedestrian track flow multistage clustering algorithm considering the multi-camera information comprises the following steps of 1. Firstly, carrying out stream space conversion on a video target data set to obtain geographic stream sub-segment sets of different levels; 2. Semi-supervised clustering is carried out on the stream sub-segment sets of different levels respectively, so that a plurality of groups of cluster center sets are obtained. The video object of step 1 is represented in the stream space: A video track OD stream consists of a series of stream subsections, a video object Obj being represented in the stream space as: fObj(i)＝{fi,1,fi,2,fi,3,