CN103324913B - A kind of pedestrian event detection method of Shape-based interpolation characteristic sum trajectory analysis - Google Patents

Abstract

The present invention provides a pedestrian event detection method based on shape features and trajectory analysis. The foreground target is obtained by target segmentation using the background difference method. The connected domain of the same target is marked with a block-based method, and the circumscribed area of the connected domain is recorded at the same time. Rectangle and extract its geometric shape features to complete the target recognition. When the pedestrian-like target is identified, the corner point of the target is extracted, and the corner point position information is used to track and match the corner points. Repeat the above process to obtain the target’s trajectory. Calculate the trajectory Segmented inflection point, linear analysis is performed in each segment formed by the inflection point to realize the calculation of the target speed, and on this basis, the status information of pedestrian events is analyzed to complete the traffic safety warning. The detection method of the invention is suitable for complex and changeable traffic scenes, can accurately identify, track and warn pedestrians appearing in the monitoring video range, has high practical value, and has broad application prospects.

Description

A Pedestrian Incident Detection Method Based on Shape Feature and Trajectory Analysis

technical field

The invention belongs to the field of video detection, and in particular relates to a pedestrian event detection method based on shape feature and trajectory analysis.

Background technique

With the development of road traffic construction, the contradiction between pedestrians and vehicles has become more and more prominent, resulting in frequent occurrence of traffic accidents. Pedestrian violations are an important cause of traffic accidents, such as running red lights, crossing the road, breaking into expressways, etc. Therefore, the monitoring of pedestrian violations has become an important part of traffic monitoring. The current traffic monitoring is mainly realized by manually checking the monitoring video and road inspection. This method is inefficient, cannot achieve real-time monitoring, and causes a great waste of resources. In the field of intelligent transportation, traditional pedestrian detection methods mainly include temperature detection, induction coil detection, sound detection, etc. Temperature detection is susceptible to interference from many heat sources in traffic scenes, resulting in false detections. The detection sensitivity of the induction coil is not high, the installation is inconvenient, and the maintainability is poor. The detection accuracy of sound detection is not high due to the noise in the traffic scene.

In recent years, video-based detection technology has been widely used due to its large detection range, real-time performance, and a lot of auxiliary information. However, due to the complexity of the traffic scene, the background and moving objects are easy to change due to factors such as light and weather. Although there are many pedestrian detection methods, such as methods based on human body parameter models and methods based on local characteristics of the human body, pedestrian event alarms can be realized. However, it cannot meet the requirements of good adaptability to environmental factors and obtaining real-time and accurate traffic detection information.

Contents of the invention

In view of the deficiencies and defects of the prior art, the object of the present invention is to provide a pedestrian event detection method based on shape features and trajectory analysis, which can overcome the complex and changeable factors of traffic road scenes, and detect pedestrian events in the detection area. Real-time and accurate detection and early warning of its danger level.

In order to realize above-mentioned task, the present invention adopts following technical scheme to realize:

A pedestrian event detection method based on shape features and trajectory analysis, the method is carried out according to the following steps:

Step 1: Establish the mapping relationship between the image pixels and the actual distance of the road surface, that is, the mapping table, and divide the road image into two parts: the road interior and the road shoulder;

Step 2: Divide the first frame image and the background image into multiple block areas in the same block coordinate system, the size of the background is W*H, the size of the divided blocks is w*h, and the number of divided block areas is T For T=(W/w)*(H/h), subtract the corresponding pixels of the current first frame image and the background image to obtain a frame difference image with a size of W*H, and divide the frame difference image into T sizes For all blocks of w*h, record the number of pixels greater than the gray threshold A in the jth block as N _j , if N _j is greater than the threshold B, then assign the value of all pixels in this block to 255, otherwise assign it to 0, in:

W is the number of pixels in the horizontal direction of the image;

H is the number of pixels in the vertical direction of the image;

w is the pixel width of the block;

h is the pixel height of the block;

j=1,2,3...T;

The value of the threshold A is 30;

The value range of the threshold B is 0.5 to 0.75 times the total number of pixels in the block;

Step 3: scan the binarized image from left to right and from top to bottom in units of blocks, mark the connected domain of the same target with the same label, and obtain the smallest circumscribing rectangle of the connected domain at the same time, and calculate the circumscribing rectangle The height R _h , width R _w , aspect ratio R _a and rectangularity R _j of the rectangle, when the value of R _a is within the range of threshold C, and the value of R _j is within the range of threshold D, keep the target, when R When _a or R _j is not within the range of thresholds C and D, remove the target, where:

The threshold C ranges from 1.5 to 8;

The threshold D ranges from 0.5 to 1;

Step 4: Find the best corner point for the jth foreground object marked on the first frame of the image, select a certain pixel point P _i (m,n) of the object as the center, establish a window with a size of a*a, and calculate The sum of the squares of the gray level differences of adjacent pixels in the horizontal, vertical and two diagonal directions passing through the center pixel P _i (m,n) is taken as the minimum value g _min in the result. If g _min is greater than the threshold E, Then this point is a corner point, if g _min is less than or equal to the threshold E, then this point is not a corner point and discarded, where:

The a is the pixel width of the side length of the window;

The value range of the threshold E is 180-220;

Step 5, record the position information of the corner point and the number of matching tracking times in a newly created empty structure, and initialize the number of matching tracking times of the target to zero;

Step 6, for the second frame, the third frame, ..., the i-th frame image, repeat the method of step 2, step 3 and step 4 to obtain the corner position of the target in the current frame, and then the corner position recorded in the previous frame is According to the comparison with the corner position of the recorded target in the current frame, there are:

When the absolute value difference between the two positions is greater than the threshold F, it is determined that the target where the corner point is located in the current frame is a new target in the current frame, and then process according to the method of step five;

When the absolute value difference between the two positions is less than or equal to the threshold F, replace the corner position information of the previous frame with the corner position information of the current frame as a new basis for comparison, and add 1 to the number of matching tracking, where:

i is a positive integer;

The value range of the threshold F is 1-5;

Step 7, when the number of matches is greater than or equal to the threshold G, the tracking is completed, and the tracking track is obtained as: Track={P _i ,P _i+1 ,...P _i+n }, and step 8 is performed, wherein:

The value range of the threshold G is 50-70;

Step eight, look up the mapping table to obtain the actual distance corresponding to each corner point in the track Track={P _i ,P _i+1 ,...P _i+n }, that is, the actual motion track Track ^, ={(s _i , 0),(s _i+1 ,1),...,(s _i+n ,n)}, where:

s _i+n represents the actual distance corresponding to the pixel point P _i+n , and n represents the subscript of the point;

Step 9, obtain the linear equation of a straight line passing through these two points by the first point P _i of the actual motion track curve and the tail point P _i+n : y=kx+b, the distance from any point on this track to this straight line is:

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; k</span>}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; |</span>}{\sqrt{{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}$

In the formula: k is the slope of the straight line, b is the intercept, (x, y) means any point on the line, (s _i+r , r) means any point on the track, d _r means (s _i+r ,r) the distance from the point to the straight line;

Sort all d _r to find the largest d _max , if it is greater than the threshold H, then this point is the inflection point of the target trajectory, save this point (s _i+r , r), and perform step ten, where:

The value of the threshold H is 70cm;

Step ten, the inflection point (s _i+r , r) divides the motion track curve into segments, the curve segment with (s ₀ , 0) and (s _i+r , r) as the first and last point, and the point (s _i+r ,r) and (s _i+n ,n) are the curve segments of the first and last points, respectively execute step 9 in these two curve segments, and continue to find the inflection points of each segment of the trajectory until the distance between all points on this trajectory and the straight line d _r ≤ H, so get a set of inflection points {(s _i+r0 ,r ₀ ),(s _i+r1 ,r ₁ ),...,(s _i+rm ,r _m )};

Step eleven, the inflection point divides the motion trajectory into motion trajectory curve segments, and uses the least square method to perform linear fitting on each motion trajectory curve segment to obtain the correlation coefficient r, then:

When r ≥ 0.5, keep the segment of the trajectory curve segment;

When r<0.5, remove the segment of the motion trajectory curve segment;

Finally, a set of motion trajectory curve segments is obtained;

Step 12, use the actual distance and time difference between the first point and the end point of each motion trajectory curve segment obtained after step 11 screening to obtain the target velocity v in the segment, the expression is:

$\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; N\&Delta;t</span>}}$

In the formula:

N represents the interval number of track points in a section of motion track curve segment;

s _f represents the actual distance of the end point of a section of motion trajectory curve segment;

s _s represents the actual distance of the first point of a section of motion trajectory curve segment;

Δt represents the time interval between two adjacent trajectory points in a segment of motion trajectory;

When the speed in all segments satisfies 0.3m/s<v<2m/s, the target can be determined to be a pedestrian;

Step 13, judge the danger level of the pedestrian event according to the position of the coordinate point P _i+n of the target in the current frame:

(1) When the point P _i+n is inside the road, the danger level of the pedestrian event is high;

(2) When the point P _{i+n is} on the shoulder, if the angle between the direction of the pedestrian trajectory vector and the correct driving direction of the road is greater than 30 degrees, the hazard level of the pedestrian incident is medium;

(3) When the point P _{i+n is} on the road shoulder, if the angle between the direction of the pedestrian trajectory vector and the correct driving direction of the road is less than or equal to 30 degrees, the hazard level of the pedestrian incident is low.

Compared with the prior art, the pedestrian event detection method based on the shape feature and trajectory analysis of the present invention is aimed at the characteristics of difficult target segmentation caused by the long sight distance of the traffic scene monitoring video and the variability of pedestrian postures, combined with the background difference The method uses the pedestrian's geometric shape features to complete target segmentation and preliminary recognition, and extracts a stable and single corner point for the target, and uses a matching algorithm to obtain the target tracking trajectory. Due to the arbitrariness of pedestrian movement, the obtained pedestrian trajectory has nonlinear characteristics. If the entire trajectory curve is directly fitted linearly, there will be a large error. Therefore, the present invention uses a segmented method to find the inflection point for the target trajectory curve. These inflection points can divide the trajectory curve into several line segments with a linear relationship, and then use the least square method to perform linear fitting on the line segments between each adjacent inflection point. After this kind of processing, more accurate target speed information can be obtained, and the accuracy rate of pedestrian event detection can be improved. In addition, the present invention can realize the judgment of the danger level of pedestrian incidents and complete the early warning function by calculating and analyzing pedestrian position information and movement direction.

Description of drawings

Figure 1 is the first frame image.

Figure 2 is the division of different areas of the road, and the white color represents the inside of the road.

Figure 3 shows the pedestrians identified based on shape features, and the white box in the figure is the bounding rectangle of the target.

Figure 4 is a schematic diagram of target feature points, where the white dots are feature corner points of the target.

Figure 5 shows the template matching search method. The black dot in the image where the corner point is located represents the feature corner point, the small square outside the black dot is the template, and the shaded part in the image to be searched is the search area. In the search area, use the template to traverse the search area and find The matching block with the minimum MAD value is used as a new template, and its center is a new corner point.

Fig. 6 is a schematic diagram of the target tracking trajectory in the video, where the white line represents the movement trajectory of the pedestrian's characteristic corners in the image.

Figure 7 is a diagram of the actual movement trajectory of the target. The position at each moment is represented by a white dot. The abscissa in the figure is time, and the unit is 0.04s; the ordinate is the actual distance, and the unit is cm.

Figure 8 is to find segmental inflection points for the actual trajectory curve, and the inflection points are marked with gray crosses.

Fig. 9 is an early warning of the danger level of the pedestrian event, showing the position and direction of movement of the pedestrian, and judging the danger level of the pedestrian event.

The content of the present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments.

detailed description

This embodiment provides a pedestrian event detection method based on shape features and trajectory analysis, through the use of background difference method for target segmentation, block-based connected domain marking, target recognition based on geometric shape features, corner point extraction, and target trajectory tracking As well as finding the segmental inflection point of the trajectory, linear analysis to achieve the target speed, and on this basis, the analysis of the status of pedestrian events is completed, and the traffic safety warning is completed.

It should be noted that the images processed in the method of the present invention are the first frame image, the second frame image, the third frame image, ..., the i (i is a positive integer) frame along the positive time sequence in the video image.

It should be noted that the mapping table in this embodiment is obtained by using the camera geometric calibration method described in the invention patent "a camera geometric calibration method under linear model" (publication (publication) number: CN102222332A).

Suppose the size of each frame of video image is W*H, and the area size of each block is w*h, wherein W is the pixel in the horizontal direction of each frame of video image, H is the pixel in the vertical direction of each frame of video image, and w is the width of each block area, and h is the height of each block area.

The method of this embodiment specifically adopts the following steps to realize:

Step 1: Establish the mapping relationship between the image pixels and the actual distance of the road surface, that is, the mapping table, and divide the road image into two parts: the road interior and the road shoulder;

Step 2: Divide the first frame image and the background image into multiple block areas in the same block coordinate system, the size of the background is W*H, the size of the divided blocks is w*h, and the number of divided block areas is T For T=(W/w)*(H/h), subtract the corresponding pixels of the current first frame image and the background image to obtain a frame difference image with a size of W*H, and divide the frame difference image into T sizes For all blocks of w*h, record the number of pixels greater than the gray threshold A in the jth block as N _j , if N _j is greater than the threshold B, then assign the value of all pixels in this block to 255, otherwise assign it to 0, in:

W is the number of pixels in the horizontal direction of the image;

H is the number of pixels in the vertical direction of the image;

w is the pixel width of the block;

h is the pixel height of the block;

j=1,2,3...T;

The value of the threshold A is 30;

The value range of the threshold B is 0.5 to 0.75 times the total number of pixels in the block;

Step 3: scan the binarized image from left to right and from top to bottom in units of blocks, mark the connected domain of the same object with the same label, and obtain the smallest circumscribing rectangle of the connected domain at the same time, and calculate the circumscribing rectangle The height R _h , width R _w , aspect ratio R _a and rectangularity R _j of the rectangle, when the value of R _a is within the range of threshold C, and the value of R _j is within the range of threshold D, keep the target, when R When _a or R _j is not within the range of thresholds C and D, remove the target, where:

The threshold C ranges from 1.5 to 8;

The threshold D ranges from 0.5 to 1;

Step 4: Find the best corner point for the jth foreground object marked on the first frame of the image, select a certain pixel point P _i (m,n) of the object as the center, establish a window with a size of a*a, and calculate The sum of the squares of the gray level differences of adjacent pixels in the horizontal, vertical and two diagonal directions passing through the center pixel P _i (m,n) is taken as the minimum value g _min in the result. If g _min is greater than the threshold E, Then this point is a corner point, if g _min is less than or equal to the threshold E, then this point is not a corner point and discarded, where:

The a is the pixel width of the side length of the window;

The value range of the threshold E is 180-220;

Step 5, record the position information of the corner point and the number of matching tracking times in a newly created empty structure, and initialize the number of matching tracking times of the target to zero;

Step 6, for the second frame, the third frame, ..., the i-th frame image, repeat the method of step 2, step 3 and step 4 to obtain the corner position of the target in the current frame, and then the corner position recorded in the previous frame is According to the comparison with the corner position of the recorded target in the current frame, there are:

When the absolute value difference between the two positions is greater than the threshold F, it is determined that the target where the corner point is located in the current frame is a new target in the current frame, and then process according to the method of step five;

When the absolute value difference between the two positions is less than or equal to the threshold F, replace the corner position information of the previous frame with the corner position information of the current frame as a new basis for comparison, and add 1 to the number of matching tracking, where:

i is a positive integer;

The value range of the threshold F is 1-5;

Step 7, when the number of matches is greater than or equal to the threshold G, the tracking is completed, and the tracking track is obtained as: Track={P _i ,P _i+1 ,...P _i+n }, and step 8 is performed, wherein:

The value range of the threshold G is 50-70;

Step eight, look up the mapping table to obtain the actual distance corresponding to each corner point in the track Track={P _i ,P _i+1 ,...P _i+n }, that is, the actual motion track Track'={(s _i , 0),(s _i+1 ,1),...,(s _i+n ,n)}, where:

s _i+n represents the actual distance corresponding to the pixel point P _i+n , and n represents the subscript of the point;

Step 9, obtain the linear equation of a straight line passing through these two points by the first point P _i of the actual motion track curve and the tail point P _i+n : y=kx+b, the distance from any point on this track to this straight line is:

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; k</span>}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; |</span>}{\sqrt{{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}$

In the formula: k is the slope of the straight line, b is the intercept, (x, y) means any point on the line, (s _i+r , r) means any point on the track, d _r means (s _i+r ,r) the distance from the point to the straight line;

Sort all d _r to find the largest d _max , if it is greater than the threshold H, then this point is the inflection point of the target trajectory, save this point (s _i+r , r), and perform step ten, where:

The value of the threshold H is 70cm;

Step ten, the inflection point (s _i+r , r) divides the motion track curve into segments, the curve segment with (s ₀ , 0) and (s _i+r , r) as the first and last point, and the point (s _i+r ,r) and (s _i+n ,n) are the curve segments of the first and last points, respectively execute step 9 in these two curve segments, and continue to find the inflection points of each segment of the trajectory until the distance between all points on this trajectory and the straight line d _r ≤ H, so get a set of inflection points {(s _i+r0 ,r ₀ ),(s _i+r1 ,r ₁ ),...,(s _i+rm ,r _m )};

Step eleven, the inflection point divides the motion trajectory into motion trajectory curve segments, and uses the least square method to perform linear fitting on each motion trajectory curve segment to obtain the correlation coefficient r, then:

When r ≥ 0.5, keep the segment of the trajectory curve segment;

When r<0.5, remove the segment of the motion trajectory curve segment;

Finally, a set of motion trajectory curve segments is obtained;

Step 12, using the actual distance and time difference between the first point and the end point of each motion trajectory curve segment obtained after step 11 screening to obtain the target velocity v in the segment, the expression is:

$\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; N\&Delta;t</span>}}$

In the formula:

N represents the interval number of track points in a section of motion track curve segment;

s _f represents the actual distance of the end point of a section of motion trajectory curve segment;

s _s represents the actual distance of the first point of a section of motion trajectory curve segment;

Δt represents the time interval between two adjacent trajectory points in a segment of motion trajectory;

When the speed in all segments satisfies 0.3m/s<v<2m/s, the target can be determined to be a pedestrian;

Step 13, judge the danger level of the pedestrian event according to the position of the coordinate point P _i+n of the target in the current frame:

(2) When the point P _i+n is inside the road, the hazard level of the pedestrian event is high;

(2) When the point P _{i+n is} on the shoulder, if the angle between the direction of the pedestrian trajectory vector and the correct driving direction of the road is greater than 30 degrees, the hazard level of the pedestrian incident is medium;

(3) When the point P _{i+n is} on the road shoulder, if the angle between the direction of the pedestrian trajectory vector and the correct driving direction of the road is less than or equal to 30 degrees, the hazard level of the pedestrian incident is low.

Specific embodiments of the present invention are provided below, and it should be noted that the present invention is not limited to the following specific embodiments, and all equivalent transformations done on the basis of the technical solutions of the present application all fall within the scope of protection of the present invention.

Example:

In the processing of this embodiment, the video sampling frequency is 25 frames per second, and the image size of each frame is 720×288. The block size of the frame difference image for block processing is 8×6, and the image is divided into 90×48 block regions. , when performing the background subtraction method, the grayscale threshold A is 30, the threshold B is 36, the aspect ratio threshold C that conforms to the characteristics of pedestrians ranges from 1.5 to 8, the rectangle threshold D ranges from 0.5 to 1, and the threshold for corner points is selected The value of E is 180-220, the threshold F of the corner matching distance is 3, the threshold G of the number of corner matching times is 50, and the threshold H of the judgment distance when looking for the segmental inflection point of the actual motion track is 70cm, as shown in Figure 1 to As shown in FIG. 9 , the video image is processed sequentially from the first frame by using the above method according to the above method.

It can be seen from Figure 6 that the white line in the figure is the pedestrian trajectory. When the video image runs to the 51st frame, the number of corner point matching reaches 50 times, so the trajectory line ends from the 1st frame to the 51st frame. The lower end of the trajectory is the position of the feature point found when the pedestrian enters the scene for the first time, and the uppermost end point is the feature point found in the 50th frame.

Figure 7 is the actual distance curve corresponding to the target tracking trajectory. Use the methods in steps 9 and 10 to obtain the inflection point of the trajectory curve. The result is shown in Figure 8, and the inflection point is marked with a cross. Then, the least square method is used to fit the trajectory in each segment, and the actual moving speed of the pedestrian can be obtained as 0.71m/s, so it can be judged that the target is a pedestrian. At this time, according to the position and direction of the pedestrian, the danger level caused by the pedestrian incident to traffic safety is judged, and traffic safety early warning is realized.

Claims (1)

1. a pedestrian event detection method for Shape-based interpolation characteristic sum trajectory analysis, is characterized in that, the method is carried out according to following steps:

Step one, sets up the mapping relations of image pixel to road surface actual range, i.e. mapping table, is divided into by road image in road and curb two parts simultaneously;

Step 2,1st two field picture is all divided into multiple pieces of regions with background image under identical block coordinate system, the size of background is W*H, the block size divided is w*h, the block areal T divided is T=(W/w) * (H/h), subtracts each other, obtain the frame difference image that size is all W*H with current 1st two field picture and background image respective pixel, frame difference image is divided into the block that T size is all w*h, the number of pixels being greater than gray threshold A in a note jth block is N _jif, N _jbe greater than threshold value B, then in this block, all pixel values are composed is 255, otherwise tax is 0, wherein:

W is image level direction number of pixels;

H is image vertical direction number of pixels;

W is the pixel wide of block;

H is the pixels tall of block;

j＝1,2,3...T；

The value of described threshold value A is 30;

The span of described threshold value B is 0.5 ~ 0.75 times of sum of all pixels in block;

Step 3, scans in units of block to binary image from left to right from top to bottom successively, is marked with identical label to the connected domain of same target, obtains the minimum enclosed rectangle of this connected domain simultaneously, calculates the height R of this boundary rectangle _h, width R _w, depth-width ratio R _awith rectangular degree R _j, work as R _avalue within the scope of threshold value C, and R _jvalue within the scope of threshold value D time, retain this target, work as R _aor R _jtime not within the scope of threshold value C and D, remove this target, wherein:

Described threshold value C scope is 1.5 ~ 8;

Described threshold value D scope is 0.5 ~ 1;

Step 4, finds best angle point to the jth foreground target that the 1st two field picture marks, and chooses this target pixel P _icentered by (m, n), setting up a size is the window of a*a, respectively calculated central pixel point P _iin horizontal, longitudinal and two diagonals of (m, n), the quadratic sum of neighbor gray scale difference, gets the minimum value g in its result _minif, the g of certain point in window _minbe greater than threshold value E, then this point in window is angle point, if g _minbe less than or equal to threshold value E, then this point is not angle point and casts out, wherein:

Described a is the pixel wide of the window length of side;

The span of described threshold value E is 180 ~ 220;

Step 5, the positional information of angle point and matched jamming number information are recorded in the structure of a newly-built sky, object matching is followed the tracks of number of times and is initialized as zero;

Step 6, to the second frame, the 3rd frame ..., the i-th two field picture, the method for step 2, step 3 and step 4 of repetition obtains the corner location of target in present frame, then with the corner location of last frame recording for foundation, compare with the corner location of the target of the record in present frame, then have:

When both positions, absolute value difference is greater than threshold value F, just determines that the target at this angle point place in present frame is target new in present frame, then process according to the method for step 5;

When both positions, absolute value difference is less than or equal to threshold value F, then replace the corner location information of former frame as new benchmark foundation by the corner location information of present frame, matched jamming number of times adds 1, wherein:

I is positive integer;

The span of described threshold value F is 1 ~ 5;

Step 7, when matching times is more than or equal to threshold value G, then follows the tracks of complete, obtains pursuit path and is: Track={P _i, P _i+1... P _i+n, perform step 8, wherein:

The span of described threshold value G is 50 ~ 70;

Step 8, searches mapping table, obtains track Track={P _i, P _i+1... P _i+nin actual range corresponding to each angle point, i.e. actual motion track Track '={ (s _i, 0), (s _i+1, 1) ..., (s _i+n, n) }, wherein:

S _i+nrepresent pixel P _i+ncorresponding actual range, n represents subscript a little;

Step 9, by actual motion geometric locus first point P _iwith tail point P _i+nobtain the straight-line equation through this straight line of 2: y=kx+b, on this track, any point to the distance of this straight line is:

$d_r=\frac{|{\mathrm{kS}}_{i+r}-r+b|}{\sqrt{k^2+1}}$

In formula: k is the slope of straight line, b is intercept, and (x, y) represents any point on this straight line, (s _i+r, r) represent any point on this track, d _rrepresent (s _i+r, r) to the distance of straight line;

To all d _rcarrying out sorts finds out maximum d _maxif be greater than threshold value H, then this point on actual motion geometric locus is the flex point of target trajectory, preserves this point (s _i+r, r), perform step 10, wherein:

The value of described threshold value H is 70cm;

Step 10, flex point (s _i+r, r) by path curves segmentation, with (s ₀, 0) and (s _i+r, be r) segment of curve that head and the tail are put, and with point (s _i+r, r) with (s _i+n, be n) segment of curve that head and the tail are put, in these two segment of curve, perform step 9 respectively, continue the flex point asking for each section of track, until all distance between beeline and dot d on this track _rtill during≤H, H is threshold value, and the value of described threshold value H is 70cm, obtains one group of flex point { (s like this _i+r0, r ₀), (s _i+r1, r ₁) ..., (s _i+rm, r _m);

Step 11, movement locus is divided into path curves section by flex point, uses least square method to carry out linear fit obtain correlation coefficient r, then have each section of path curves section:

When r >=0.5, retain this section of path curves section;

As r < 0.5, remove this section of path curves section;

Finally obtain one group of path curves section;

Step 12, utilize the first point of each path curves section that obtains and the actual range of tail point and mistiming after step 11 screening to ask for segmentation internal object speed v, expression formula is:

$v=\frac{|s_f-s_s|}{N\mathrm{\&Delta;}t}$

In formula:

N represents the interval hop count of tracing point in one section of path curves section;

S _frepresent the tail point actual range of one section of path curves section;

S _srepresent the first point actual range of one section of path curves section;

Δ t represents the time interval of adjacent two tracing points in one section of path curves section;

When the speed in all segmentations all meets 0.3m/s < v < 2m/s, can determine that this target is for pedestrian;

Step 13, according to the coordinate points P of target at present frame _i+nposition judgment pedestrian event danger classes:

(1) as a P _i+nwhen being in road inside, this pedestrian's event danger classes is high;

(2) as a P _i+nwhen being in curb, if the direction of pedestrian movement's track vector and road correctly exercise angular separation be greater than 30 degree, then during this pedestrian's event danger classes is;

(3) as a P _i+nwhen being in curb, if the direction of pedestrian movement's track vector and road correctly exercise angular separation be less than or equal to 30 degree, then this pedestrian's event danger classes is low.