CN111260696B

CN111260696B - Method for edge-end-oriented pedestrian tracking and accurate people counting

Info

Publication number: CN111260696B
Application number: CN202010068506.1A
Authority: CN
Inventors: 黄樟钦; 盛梦雪; 张硕; 李洪亮
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2020-01-21
Filing date: 2020-01-21
Publication date: 2023-04-18
Anticipated expiration: 2040-01-21
Also published as: CN111260696A; WO2021147826A1

Abstract

An edge-end-oriented pedestrian tracking and people number accurate counting method relates to the field of pedestrian tracking and people number counting based on artificial intelligence components in an edge calculation scene. The method initially establishes and maintains a pedestrian dynamic tracking list in real time. And acquiring pedestrian result information from the artificial intelligence component at the edge end, and establishing a pedestrian static detection list on the basis of the pedestrian result information. And accurately matching the pedestrian static detection list with the pedestrian dynamic tracking list. Cost and performance limitations at the edge end make the edge end computing power limited, leading to failure of pedestrian exact matching. At the moment, the method carries out intelligent fuzzy matching on the pedestrian according to the dynamic tracking information of the pedestrian. And updating a pedestrian dynamic tracking list in real time according to the matched result, accurately counting the number of the pedestrians in the counting area, and if the accurate counting fails, using a delay counting strategy for the pedestrians by the method.

Description

Method for edge-end-oriented pedestrian tracking and accurate people counting

Technical Field

The invention relates to the field of pedestrian tracking and people counting based on artificial intelligence components in an edge calculation scene.

Background

With the general improvement of safety consciousness of people, monitoring cameras are distributed all over streets and alleys, and the integration of technologies such as image processing, machine vision and the like enables a computer to replace human brain to analyze information in videos, judge the conditions in the videos and enable monitoring to be intelligent. In recent years, edge computing has emerged, and embedded systems become the most important processing platform of edge computing due to the characteristics of low power consumption and single function, so that intelligent monitoring technology can be widely applied to actual life.

The efficiency of the embedded special chip is far higher than that of a PC (personal computer), the cost and the power consumption are greatly lower than those of a CPU (central processing unit), and a plurality of scholars design the special chip for embedded image processing and design methods for pedestrian tracking and people counting in a targeted manner. The pedestrian tracking strategy comprises the steps of carrying out similarity matching by utilizing the characteristic information of color, posture and the like of pedestrians analyzed by a special chip or judging the overlapping rate of front and rear frame detection frames. People counting is mainly triggered by crossing detection lines.

Artificial intelligence technology is now rapidly falling to the ground, and is also expanding towards the edge and terminal areas. In recent years, each large designer specially designs an artificial intelligence component for an artificial intelligence algorithm, however, the requirement of the artificial intelligence component on the computing capacity of a deployment platform is high, the processing capacity of an edge computing platform may limit the processing performance of the artificial intelligence component, and meanwhile, each large designer also tends to reduce the processing performance of the artificial intelligence component to meet the balance of cost and selling price in practical application.

Disclosure of Invention

In view of the above, the objective of the algorithm is to design a general pedestrian tracking and people counting method based on pedestrian result information obtained by intelligently analyzing a real-time video and related information thereof by an artificial intelligence processing component, so that when the algorithm is applied to an edge computing scene, the pedestrian tracking and people counting can still be accurately performed under the resource limitations of insufficient processing capability of the artificial intelligence processing component, limited computing capability of an edge end processing platform, product cost pricing and the like.

A method for pedestrian tracking and people number accurate statistics facing to an edge end is characterized by comprising the following steps:

step S1: and establishing and maintaining a pedestrian dynamic tracking list in real time for dynamically recording pedestrian information in a tracking area.

Step S2: and acquiring pedestrian result information of the edge artificial intelligence processing component.

And step S3: and extracting important pedestrian result information to establish a pedestrian static detection list.

And step S4: and matching the pedestrians in the static pedestrian detection list and the dynamic pedestrian tracking list.

Step S5: and updating the pedestrian dynamic tracking list according to the matching result.

Step S6: and accurately counting the number of the pedestrians in the counting area according to the dynamic pedestrian tracking list.

The information recorded in the pedestrian dynamic tracking list in the step S1 should include the tracking id, direction, moving radius, and frame loss information of the pedestrian.

The information recorded in the static pedestrian detection list in step S3 should include the tracking id, position coordinate, and time information of the pedestrian.

The algorithm is further characterized by: the resource limitation of the edge end, including the self processing capacity limitation of the artificial intelligent processing part, the insufficient computing capacity of the edge end processing platform, the product cost limitation and the like, can cause the intelligent analysis process of the real-time video to have the frame loss problem, thereby causing the deviation of the pedestrian result information. Therefore, the matching process in the step S4 is firstly carried out with accurate matching, and if the accurate matching fails, the pedestrian with the failed accurate matching is carried out with intelligent fuzzy matching.

Step S41: and (4) precise matching.

And judging whether the pedestrians in the two lists are the same person or not according to the tracking id in the static detection list and the dynamic tracking list of the pedestrians, if so, judging that the pedestrians in the two lists are the same person, and if not, judging that the pedestrians in the two lists are the same person, so that the accurate matching is successful, otherwise, the accurate matching is failed.

Step S42: and carrying out intelligent fuzzy matching on the pedestrians with the accurate matching failure.

For the pedestrians which do not have the frame loss problem and have definite directions, the position range which is possibly generated in the next frame is predicted in the motion direction of the pedestrians, and the pedestrians which have the largest similarity with the prediction result and exceed the matching threshold MATCH _ THRESHOID in the static detection list are selected to be matched with the pedestrians. If MATCH _ THRESHOID is percentage, then MATCH _ THRESHOID is more than or equal to 0 and less than or equal to 100 percent, and the matching condition is more severe as the MATCH _ THRESHOID approaches 100 percent, and generally the matching condition is set to be more than or equal to 80 percent.

And for pedestrians which do not have the frame loss problem but have ambiguous directions, bidirectionally predicting the position range of the pedestrians which may appear in the next frame, and selecting the pedestrians which have the largest similarity with the prediction result and exceed MATCH _ THRESHOID in the static detection list to be matched with the pedestrians. After matching is successful, the direction of the pedestrian can be judged according to the moving condition of the position coordinate.

For the pedestrian with frame loss problem, calculating the number n of the frame loss, and expanding the prediction range by a multiple of alpha x n, wherein alpha is an expansion coefficient, and the matching range is a part of a circle with a moving radius as a radius under the general condition, and alpha = pi n ² . When n is less than or equal to LOST _ THRESHOID, selecting static detection list and predictionThe pedestrian with the largest similarity and exceeding MATCH _ threshold is matched with the result. LOST _ THRESHOID is a frame loss threshold, and LOST _ THRESHOID =3 can be generally set.

And for the pedestrians with changed directions, the pedestrians with changed directions can not be successfully matched through the steps, the position range of the pedestrians with changed directions possibly appearing in the next frame is reversely predicted, and the pedestrians with the maximum similarity to the prediction result and exceeding MATCH _ THRESHOID in the static detection list are selected to be matched with the pedestrians. If the matching is successful, recording the direction of the sudden change.

The step S5 of updating the pedestrian dynamic tracking list includes:

and recording frame loss information when the pedestrians which are not successfully matched in the dynamic tracking list are not matched.

And (4) marking the pedestrians which are not successfully matched in the static detection list as new pedestrians, judging the direction and adding the new pedestrians into the dynamic tracking list.

And (4) walking out of the tracking area or deleting the pedestrians with the frame loss number larger than LOST _ THRESHOID, and deleting the dynamic tracking list.

The algorithm is further characterized by: the direction is used as an intelligent fuzzy matching index, so that the accuracy of people counting is improved better. And judging the direction of the pedestrian in the static detection list according to the initial position coordinate when the pedestrian is judged to be a new person. If the initial position coordinates of the pedestrian are located at the upper part of the tracking area, the initial position coordinates are generally regarded as the upper 1/3 part of the tracking area, and the direction is judged to be entering. If the initial position coordinate of the pedestrian is located at the lower part of the tracking area, the initial position coordinate is generally regarded as the lower 1/3 part of the tracking area, and the direction is judged to be out. If the initial position coordinates of the pedestrian are located in other parts of the tracking area, it is considered that the direction cannot be judged, and the direction can be confirmed with delay according to the processing method of the pedestrian whose frame loss problem does not occur but the direction is unknown in step S42, so as to make a correct judgment on the direction.

The accurate people counting process in the step S6 comprises the following steps:

step S61: and establishing and maintaining a pedestrian dynamic statistical column, and using the table for accurate statistics of the number of people.

Step S62: the accurate statistics of the number of the pedestrians in the statistical area is specifically as follows:

if the pedestrian direction is known and enters the statistical area for the first time, the total number of people and the total number of people in the direction are recorded.

If the direction of the pedestrian is unknown and the pedestrian enters the counting area for the first time, only the total number of people is recorded, and an intelligent delay counting strategy is adopted for the total number of people in the direction.

If the direction of the pedestrians changes suddenly and enters the statistical area for the first time, the total number of people is increased, and the counting with the direction is not carried out.

Step S63: when the pedestrian is counted normally and leaves the statistical area or is in the statistical area but the number of dropped frames is greater than LOST _ THRESHOID, it is deleted from the pedestrian dynamic statistical list.

The accurate counting of the number of the pedestrians cannot be carried out when the dynamic tracking information of the pedestrians is ambiguous, and the algorithm carries out intelligent delay counting processing on the pedestrians, and specifically comprises the following steps:

if the pedestrian enters the statistical area but the direction is not determined, counting is not carried out until the direction is determined through the pedestrian tracking process.

If the pedestrian is not counted normally but leaves the statistical area, the pedestrian is not deleted until the direction is determined through the pedestrian tracking process and the pedestrian is counted normally.

If the pedestrian is not counted normally but the number of frame loss is larger than LOST _ THRESHOID, the pedestrian is deleted. Considering the actual situation of the set statistical region and LOST _ THRESHOID, if the pedestrian is considered not to be detected again in the statistical region, the pedestrian is counted in the total number, otherwise, the pedestrian is ignored.

In the algorithm, except for an intelligent delay counting strategy, counting is only carried out when a pedestrian enters a statistical area for the first time, and when the pedestrian stays in the statistical area, the counting is not repeated although the pedestrian stays in a dynamic statistical list of the pedestrian.

The algorithm of claim 1, further characterized by: a method for delineating a range of positions where a pedestrian is likely to be present in a next frame. The scribing process is detailed as follows:

one of the most rough methods is to define the prediction range within a full circle with the center of mass of the pedestrian as the center and the radius of movement as the radius.

Further, the method uses the direction as an important matching index, and the limitation of the direction helps us to reduce the prediction range from a full circle to a semicircle in the moving direction. The semicircle is specifically: taking the center of mass of the pedestrian as an end point, making a ray parallel to the moving direction, dividing a whole circle into semicircles by a straight line which passes through the end point and is perpendicular to the ray, and taking the semicircle containing the ray.

Furthermore, the time difference between the two video frames before and after being analyzed intelligently is always within 1 second, and the moving direction of the pedestrian in 1 second can not be changed greatly, specifically, the angle of the pedestrian to the left or right in the moving direction can not be greater than 45 degrees, so that the prediction range is reduced to the sector of the semi-circle with the radius and the included angle of the ray being less than 45 degrees and > -45 degrees.

The algorithm has the advantages that:

(1) The method is universal for various detection models such as face detection, pedestrian detection, head and shoulder detection and the like, the accurate matching conditions can be changed according to actual conditions, unidirectional or bidirectional people counting can be carried out, and the method has good expandability and universality.

(2) The intelligent fuzzy matching method has the advantages that the problem that the accuracy of people counting is reduced due to errors of pedestrian result information caused by the problems of limited processing capacity of an artificial intelligent processing part, insufficient computing capacity of an edge end processing platform, product cost pricing and other resource limitations is considered, and the intelligent fuzzy matching is added on the basis of accurate matching, so that the pedestrian tracking and the people counting are more accurate.

Description of the drawings:

FIG. 1 Algorithm flow chart

FIG. 2 is a diagram illustrating a rough prediction range

FIG. 3 is a schematic view of the prediction range of the tape direction

FIG. 4 is a schematic diagram of the prediction range with angles

FIG. 5 is a diagram of prediction range with matching distance coefficients

FIG. 6 is a diagram illustrating bi-directional match prediction range

FIG. 7 is a diagram illustrating the prediction range of reverse matching

The specific implementation mode is as follows:

the technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention, and the examples are only for explaining the present invention and do not limit the application of the present invention.

In this embodiment, the pedestrian result information provided by the artificial intelligence component to us includes the tracking id, the position coordinates, and the current frame number Seq of the pedestrian _curr Face feature vectors, etc. The performance of the artificial intelligence component is good, the processing rate is 12 frames per second, but the preprocessing capability of the edge end processing platform for video frames is limited, and only 3 frames per second can be achieved, so that the analysis is performed at intervals of fixed frame number frames (which are called as detection frames ofps), in the embodiment, frame =7, the same video frames are used as input data of the artificial intelligence component among groups, which inevitably causes the processing performance of the artificial intelligence component to be reduced, and the pedestrian tracking id is inaccurate.

The invention provides a method for edge-end-oriented pedestrian tracking and accurate people number counting, which comprises the following specific steps of:

step S1: establishing and maintaining a pedestrian dynamic tracking list and a pedestrian dynamic statistical list in real time, wherein the pedestrian dynamic tracking list is used for dynamically recording pedestrian information in a tracking area, and adding pedestrians into the pedestrian dynamic statistical list when the pedestrians enter the statistical area, and the two lists comprise the following 4 characteristics:

1) Tracking id: included in the pedestrian result information provided by the artificial intelligence component.

2) Moving direction d: if the initial position coordinates of the pedestrian are located in the upper 1/3 part of the tracking area, we consider the pedestrian as the entering direction, and make it carry the entering direction attribute, note d = in. If the initial position coordinates of the pedestrian are located in the lower 1/3 part of the tracking area, we consider the pedestrian as an outgoing direction, and make it carry out the directional attribute, note d = out. If the initial position coordinates of the pedestrian are located in other parts of the tracking area, the direction cannot be judged, the direction attribute is not carried, and the direction is confirmed in a delayed mode.

3) Moving radius r: defined as the pedestrian starting from entering the tracking area until the presentThe average distance moved by each detection frame. Let the initial centroid coordinate of the pedestrian entering the tracking area be (x) _orgi ，y _orgi ) The coordinates of the centroid of the current frame are (x) _curr ，y _curr ) The initial frame number is Seq _orgi The current frame number is Seq _curr Then, then

4) Frame number lost Seq _lost : seq when no frame is lost _lost = -1, if the pedestrian is not successfully matched with any pedestrian in the pedestrian static detection list after one round of algorithm, recording frame loss, seq _lost ＝Seq _curr 。

Step S2: and acquiring pedestrian result information of the current video frame from the artificial intelligence component, establishing a pedestrian static detection list on the basis, and matching pedestrians in the static detection list with pedestrians in the dynamic tracking list.

The basic principle of the algorithm matching is as follows: firstly, carrying out accurate matching, namely, if the tracking id is the same, the accurate matching is successful, otherwise, according to an intelligent fuzzy matching method, if the pedestrian direction in the dynamic tracking list is known, selecting the pedestrian which is positioned in the moving direction of the dynamic tracking list, is in the prediction range and is closest to the dynamic tracking list to be matched with the pedestrian.

The prediction horizon of the algorithm we make the following analysis:

one of the coarsest approaches may be to define the prediction range as being the pedestrian's centroid (x) _curr ，y _curr ) As the center of the circle, the circle with the radius of movement r as the radius, as shown in fig. 2. The basic matching principle of the algorithm can help us to narrow the prediction range from fig. 2 to fig. 3 to be within a semicircle in the moving direction by adding the limit of the moving direction d. However, the algorithm can also perform bidirectional people counting, so that the situation that two pedestrians walking in the same direction get closer before passing by the shoulder (as shown by two pedestrians id2 and id1 in fig. 4) to cause matching error may occur. Therefore, it is not enough to reduce the prediction range to fig. 3, and we think that two pedestrians will avoid the collision as much as possible, so it is considered that the pedestrians are confronted with each otherWhen the walking is too close, the walking and the walking are mutually avoided to generate an angle as shown in the graph 4, the time difference between the two video frames before and after being analyzed by the intelligent analysis is certainly within 1 second, the moving direction of the pedestrian does not change greatly within 1 second, and particularly, the angle of the pedestrian deviating to the left or the right in the moving direction is not larger than 45 degrees, so that the prediction range is further reduced to the graph 4, namely, the radius is in a sector with the included angle of less than 45 degrees and > -45 degrees with the ray. In conclusion, the algorithm considers that the part outside the shadow of FIG. 4 does not belong to the prediction range (so in FIG. 4, we consider id3 to be a better intelligent fuzzy matching object with id 1).

The algorithm uses a matching table for analysis, with the transverse representing pedestrians in a dynamic tracking list and the longitudinal representing pedestrians in a static detection list. The values that may appear in the table and their meanings are given in the following table:

numerical value	Means of
		-2	Out of prediction range
-1	Has been matched
		0	Tracking id identity
distance (other value)	Relative distance between two pedestrian centroids

The form filling process is as follows:

step S21: if the two tracking ids are the same, the artificial intelligence component analyzes to obtain that the two persons are the same person, the accurate matching is finished, and 0 is filled in

Step S22: if the pedestrian in the static detection list is not within the predicted range of the pedestrian in the dynamic tracking list, fill-2.

Step S23: otherwise, distance fill is calculated.

And step S3: sequentially extracting minimum distances other than-2 from the table _min If the following relationships are satisfied simultaneously:

1) The pedestrians of the corresponding dynamic tracking list and the pedestrians of the static detection list are not matched.

2) If MIN _ COEFFICIENT is the minimum COEFFICIENT of the matching distance and MAX _ COEFFICIENT is the maximum COEFFICIENT of the matching distance, MIN _ COEFFICIENT multiplied by r is not more than distance _min MAX _ COEFFIIENT x r, the prediction horizon is further refined in this step, as shown in FIG. 5. The MIN _ COEFFICENT and MAX _ COEFFICENT can be selected to be the optimal values within the range of 0-1 according to actual conditions. If MATCH _ threshold =80%, MIN _ coeffient =0.9 and MAX _ coeffient =1.1 may be set.

We consider the two pedestrians as successfully matched and set the correspondence to-1, indicating that the two pedestrians have been matched.

Because the frame loss problem caused by the performance reduction of the artificial intelligence component influences the accuracy of pedestrian tracking and accurate people counting, the intelligent fuzzy matching is carried out on the pedestrians which are not successfully matched after the three steps. Specifically, as shown in step S4:

the delay confirmation strategy of the moving direction d in the step S1 is specifically:

the pedestrian with uncertain direction in the dynamic tracking list is subjected to two-way range matching, namely, the prediction range is expanded from the graph 5 to the graph 6, the pedestrian with the nearest distance in the range is selected to be matched with the pedestrian, and after the matching is successful, the moving direction of the pedestrian can be confirmed according to the moving direction of the center of mass of the pedestrian.

Frame loss compensation strategy:

by calculating we canObtaining the ofps of a plurality of detection frames lost by the pedestrian according to the calculation formula

Accordingly we also enlarge the matching radius of movement r by a factor of n. Namely, MINCEOEFFICIENT × r × n < distance _min < MAX _ COEFFICIENT × r × n. The pedestrian is deleted from the pedestrian dynamic tracking list when n > LOST _ THRESHOID. If framestap =7, lost _threshold =3.

Sudden change of the pedestrian direction:

if the pedestrians who are not successfully matched still exist after the judgment, the pedestrians are considered whether the moving direction is changed or not, so that the pedestrians are subjected to reverse matching as shown in the figure 7, and if the matching is successful, the pedestrians are marked to have sudden changes of the direction. When the number of people is accurately counted, the total number of people is calculated, but the direction-carrying counting is not carried out.

Step S5: and updating the pedestrian dynamic tracking list according to the accurate matching and intelligent fuzzy matching results.

The pedestrians which are not successfully matched in the dynamic tracking list are recorded with Seq _lost ＝Seq _curr 。

And (4) recording the pedestrians which are not successfully matched in the static detection list as new pedestrians, judging the moving direction d, and adding the new pedestrians into the dynamic tracking list.

Step S6: the accurate statistics of the number of the pedestrians in the statistical area is specifically as follows:

step S61: if the pedestrians enter the statistical area for the first time and d = in, the total number of people is increased, and the number of people in the direction of the in is increased.

If the pedestrians enter the statistical area for the first time and d = out, the number of the pedestrians in the total area is increased, and the number of the pedestrians in the out direction is increased.

If the pedestrians enter the statistical area for the first time and the direction is unknown, only the total number of people is recorded, and an intelligent delay counting strategy is adopted for the total number of people in the direction.

If the pedestrians enter the statistical area for the first time and carry the direction mutation attribute, the total number of the pedestrians is increased, and the pedestrians are not counted in the direction.

Step S62: when the pedestrian is counted normally and leaves the statistical area or the number of frames LOST in the statistical area is greater than LOST _ THRESHOID, the pedestrian is deleted from the dynamic statistical list.

The intelligent delay counting strategy in step S61 is specifically:

If the pedestrians do not count normally but walk out of the statistical area, the pedestrians are not deleted until the direction is determined and the pedestrians count normally in the tracking process.

If the pedestrian is not counted normally but the number of frame loss is larger than LOST _ THRESHOID, the pedestrian is deleted. Considering the actual situation of the statistical region and LOST _ THRESHOID set, if the pedestrian is not detected again in the statistical region, the pedestrian is counted in the total number, otherwise, the pedestrian is ignored.

In the algorithm, except for a delay counting strategy, counting is only carried out when a pedestrian enters a statistical area for the first time, and when the pedestrian stays in a dynamic statistical area, although the pedestrian stays in a statistical list, the counting is not repeated.

Claims

1. A method for tracking pedestrians facing to an edge end and accurately counting the number of the pedestrians is characterized by comprising the following steps of:

step S1: establishing and maintaining a pedestrian dynamic tracking list in real time for dynamically recording pedestrian information in a tracking area;

step S2: acquiring pedestrian result information of an edge artificial intelligence processing component;

and step S3: extracting important pedestrian result information and establishing a pedestrian static detection list;

and step S4: matching the pedestrians in the static pedestrian detection list and the dynamic pedestrian tracking list;

step S5: updating a pedestrian dynamic tracking list according to the matching result;

step S6: accurately counting the number of pedestrians in the counting area according to the dynamic pedestrian tracking list;

the information recorded by the pedestrian dynamic tracking list in the step S1 comprises the tracking id, direction, moving radius and frame loss information of the pedestrian;

in the step S3, the information recorded by the pedestrian static detection list comprises the tracking id, the position coordinate and the time information of the pedestrian;

therefore, the matching process in the step S4 is firstly carried out with accurate matching, and if the accurate matching fails, the pedestrian with the failed accurate matching is carried out with intelligent fuzzy matching;

step S41: precise matching;

whether the pedestrians in the two lists are the same person can be judged according to the tracking id in the static detection list and the dynamic tracking list of the pedestrians, if the tracking id is the same person, the pedestrians are the same person, the accurate matching is successful, and if not, the accurate matching is failed;

step S42: carrying out intelligent fuzzy matching on the pedestrians with the accurate matching failure;

for the pedestrians which do not have the frame loss problem and have definite directions, predicting the position range which possibly appears in the next frame in the motion direction of the pedestrians, and selecting the pedestrians which have the largest similarity with the prediction result and exceed the matching threshold MATCH _ THRESHOID in the static detection list to be matched with the pedestrians; if MATCH _ THRESHOID is percentage, then MATCH _ THRESHOID is more than or equal to 0 and less than or equal to 100 percent, and MATCH _ THRESHOID is set to be more than or equal to 80 percent;

for pedestrians who do not have the frame loss problem but have ambiguous directions, performing bidirectional prediction on the possible position range of the pedestrians in the next frame, and selecting the pedestrian which has the largest similarity with the prediction result and exceeds MATCH _ THRESHOID in the static detection list to MATCH with the pedestrian; judging the direction of the pedestrian according to the moving condition of the position coordinate of the pedestrian after successful matching;

for the pedestrian with frame loss problem, calculating the number n of the frame loss, expanding the prediction range by alpha multiplied by n, wherein alpha is an expansion coefficient, the matching range is a part of a circle with the moving radius as the radius, and alpha = pi n ² (ii) a When n is less than or equal to LOST _ THRESHOID, selecting the pedestrian with the maximum similarity to the prediction result and exceeding MATCH _ THRESHOID in the static detection list to be matched with the pedestrian; LOST _ THRESHOID is a frame loss threshold, and LOST _ THRESHOID =3;

for the pedestrians with changed directions, the pedestrians with changed directions can not be successfully matched through the steps, the position range of the pedestrians with changed directions possibly appearing in the next frame is reversely predicted, and the pedestrians with the largest similarity to the prediction result and exceeding MATCH _ THRESHOID in the static detection list are selected to be matched with the pedestrians; if the matching is successful, recording the direction of the terminal to generate mutation;

the step S5 of updating the pedestrian dynamic tracking list includes:

if the pedestrian is not successfully matched in the dynamic tracking list, recording frame loss information;

the pedestrians which are not successfully matched in the static detection list are marked as new people, the direction is judged, and a dynamic tracking list is added;

2. The method of claim 1, wherein: judging the direction of the pedestrian in the static detection list according to the initial position coordinate when the pedestrian is judged to be a new person; if the initial position coordinate of the pedestrian is positioned at the upper part of the tracking area, the initial position coordinate of the pedestrian is regarded as the upper 1/3 part of the tracking area, and the direction is judged to be entering; if the initial position coordinate of the pedestrian is positioned at the lower part of the tracking area, the initial position coordinate of the pedestrian is regarded as the lower 1/3 part of the tracking area, and the direction is judged to be out; if the initial position coordinates of the pedestrian are located in other parts of the tracking area, it is considered that the direction cannot be judged, and the direction can be confirmed with delay according to the processing method of the pedestrian whose frame loss problem does not occur but the direction is unknown in step S42, so as to make a correct judgment on the direction.

3. The method of claim 1, wherein: the accurate people counting process in the step S6 comprises the following steps:

step S61: establishing and maintaining a pedestrian dynamic statistical column, wherein a table is used for accurately counting the number of people;

if the direction of the pedestrians is known and the pedestrians enter the statistical area for the first time, recording the total number of the pedestrians and the total number of the pedestrians in the direction;

if the direction of the pedestrian is unknown and the pedestrian enters the counting area for the first time, only the total number of people is recorded, and an intelligent delay counting strategy is adopted for the total number of people in the direction;

if the pedestrian direction changes suddenly and enters the statistical area for the first time, the total number of people is increased, and the counting with the direction is not carried out;

4. The method of claim 1, wherein: the accurate statistics of the number of the pedestrians cannot be carried out when the dynamic tracking information of the pedestrians is not clear, and the algorithm carries out intelligent delay counting processing on the pedestrians, and specifically comprises the following steps:

if the pedestrian enters the statistical area but the direction is not determined, counting is not carried out until the direction is determined through the pedestrian tracking process;

if the pedestrian is not counted normally but leaves the statistical area, the pedestrian is not deleted until the direction is determined through the pedestrian tracking process and the pedestrian is counted normally;

if the pedestrian is not counted normally but the frame loss number is larger than LOST _ THRESHOID, deleting the pedestrian; considering the set statistical region and the actual situation of LOST _ THRESHOID, if the pedestrian is not detected again in the statistical region, the pedestrian is counted into the total number, otherwise, the pedestrian is ignored;

5. The method of claim 1, wherein: the method for defining the position range of the pedestrian possibly appearing in the next frame specifically comprises the following steps:

defining the prediction range in a whole circle with the center of mass of the pedestrian as the center of a circle and the moving radius as the radius;

the direction limitation reduces the prediction range from a full circle to a semicircle in the moving direction of the prediction range; the semicircle is specifically: taking the center of mass of the pedestrian as an end point, making rays parallel to the moving direction, dividing a whole circle into semicircles by a straight line which passes through the end point and is perpendicular to the rays, and taking the semicircle containing the rays;

the time difference between the two video frames is always within 1 second, and the angle of the moving direction to the left or right is not more than 45 degrees, so the prediction range is reduced to the sector of which the radius of the semicircle and the ray form an angle of <45 degrees and > -45 degrees.