CN112288767A - Automatic detection and tracking method based on target adaptive projection - Google Patents

Abstract

The invention discloses an automatic detection and tracking method based on target adaptive projection, which comprises the following steps: reading video sequence information to be tracked, and performing RGB-Gray color space conversion on each frame of image; obtaining a foreground target mask in the previous M frames by using a weighted sliding variance method; adaptively selecting a mask with a maximum foreground target area; horizontally and vertically projecting the mask; searching the projection result in the vertical direction and the horizontal direction, and determining the center of the optimal target by accumulating pixels and the occupied proportion; starting from the central position, carrying out reciprocating detection in the vertical and horizontal directions, and determining the optimal target dimension by accumulating pixels and occupied proportion so as to determine an ROI (target region); and finally, tracking the target by using a centroid method to finish the automatic target detection and tracking process. The method is simple to implement, the flow is clear, and the process of extracting the ROI by the self-adaptive projection method can effectively avoid the influence of background noise.

Description

Automatic detection and tracking method based on target adaptive projection

Technical Field

The invention belongs to the field of industrial computer vision, and relates to the problems of background modeling, target detection, target tracking and initial ROI calculation.

Background

In the field of computer vision, target detection and target tracking are two hot research problems, and the research process aiming at the two problems is slowly transited from a traditional method to a depth method. In most research works, target detection and target tracking are studied as two separate problems, and ideally, for a given video sequence, the algorithm can automatically frame the target and automatically track it. Although the deep learning method has achieved good results in recent years, the deep learning method requires a large amount of data sets to train the model, and in a real scene, the data acquisition is often difficult. Therefore, the conventional method still has its indispensable role, and in the conventional method, the method of background modeling is a very feasible solution for target detection, but the output of background modeling may contain noise, which may affect the accuracy of the extracted target ROI to some extent. Therefore, how to accurately extract the target ROI according to the foreground mask output by the background modeling method becomes a key problem.

Disclosure of Invention

The invention discloses an automatic detection and tracking method based on target self-adaptive projection, aiming at accurately extracting an ROI of a target to be detected and effectively tracking the ROI.

The invention relates to an automatic detection and tracking method based on target adaptive projection, which is characterized by comprising a target detection step near an initial frame, an adaptive mask selection step, an ROI extraction step and a target tracking step; the detection step of the target near the initial frame is a weighted sliding variance method; the self-adaptive mask selecting step is to select the optimal foreground mask; the ROI extraction step is a self-adaptive projection method; the target tracking step is a centroid method.

Preferably, reading the video sequence information to be tracked, and performing RGB-to-Gray color space conversion on each frame of image;

preferably, the step of calculating the image mask by the weighted moving average method is as follows:

step (1): for the k frame image F_kCalculating the weighted mean image thereof according to the following formula

Step (2): computing weighted variance sum S of kth frame_kAs shown in the following formula:

S_k:＝0.5V_k+0.3V_k-1+0.2V_k-2

wherein the content of the first and second substances,

and (3): calculating S_kIs squared and normalized to [0,255 ]]Interval, get

Setting a threshold value T, dividing the value according to T

And obtaining a binary mask from the pixels in the image.

Preferably, a target mask of the previous M frames is obtained by using a weighted sliding variance method, and the optimal mask is the mask with the largest proportion of the foreground area.

Preferably, the adaptive projection is performed by using an optimal mask, and the steps are as follows:

step (1): performing horizontal and vertical projection on the mask, and counting the sum of horizontal and vertical pixels, which is D_h,D_v；

Step (2): calculating the center position (P) of the optimal target_x,P_y)：

And (3): setting the threshold Q to 0.97, avoiding the influence of noise, and calculating the optimum scale (R)_w,R_h) The pseudo code is as follows:

preferably, the effective tracking is performed by using a centroid method, which comprises the following steps:

step (1): image I of the best mask obtained by weighted sliding variance method, and ROI (x, y, ROI) of the target obtained by adaptive projection method_W,roi_H) Calculating the centroid position of the current frame

Initializing a tracker;

step (2): inputting the next frame image, and converting it into grayscale image I_next；

And (3): the inter-frame difference D within the ROI is calculated using the following equation:

D(i-x,j-y):＝|I_next(i,j)-I(i,j)|,x≤i≤x+roi_W,y≤j≤y+roi_H

and (4): calculating a segmentation threshold Th by adopting an OTSU method, and binarizing D by using the Th;

and (5): finding the centroid C by_x,C_y：

And (6): the coordinates (x, y) of the ROI are updated with the following formula:

and (7): and (5) returning the updated ROI and repeating the step (2) until the video reading is finished.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is the result of adaptive projection method to calculate the target ROI (in the case of background noise);

FIG. 3 is the result of tracking by centroid method at frame 53;

fig. 4 algorithm pseudo code of the adaptive projection method.

Detailed Description

The specific implementation method is performed as the steps shown in fig. 1, first, reading the previous M frames of images, calculating a foreground mask by using a weighted sliding variance method, and obtaining an optimal mask, namely the mask with the largest proportion of foreground objects in the whole image; performing adaptive projection on the mask, calculating the center of the target, and adaptively determining the width and the height of the target by utilizing the algorithm shown in FIG. 4; the ROI is used for initializing the centroid tracker, and the optimal mask is not necessarily the Mth frame image, so that the ROI needs to be tracked and updated by a centroid method frame by frame aiming at the images from the optimal mask to the Mth frame image, so that the potential defect caused by directly tracking from the Mth frame image is prevented, namely the target of the Mth frame image can be greatly different from the target of the image where the optimal mask is located.

Results of the experiment

Fig. 2 shows a mask obtained by the weighted sliding variance method and an ROI obtained by the adaptive projection method, wherein the interference of background noise on the ROI can be effectively avoided by adjusting the threshold Q. Moreover, the background noise has small pixel sum compared with the foreground target, so the foreground target can be effectively extracted by adopting the self-adaptive projection method. Of course, although the background noise can be removed to a certain extent by using the morphological operation, the shape of the foreground object may be affected at the same time, and the extraction effect of the foreground object may be affected. Fig. 3 shows the result of tracking by centroid method at 53 frames, and it can be seen that the foreground object is stably nested by the tracking frame even if there is some deformation in the object.

Claims (6)

1. An automatic detection and tracking method based on target adaptive projection is characterized by comprising a target detection step near an initial frame, an adaptive mask selection step, an ROI extraction step and a target tracking step; the detection step of the target near the initial frame is a weighted sliding variance method; the self-adaptive mask selecting step is to select the optimal foreground mask; the ROI extraction step is a self-adaptive projection method; the target tracking step is a centroid method.

2. The automatic detection and tracking method of claim 1, wherein the video sequence information to be tracked is read and RGB to Gray color space conversion is performed for each frame of image.

3. The automatic detection and tracking method of claim 1 wherein the weighted moving average method finds the image mask by:

step (1): for the k frame image F_kCalculating the weighted mean image thereof according to the following formula

Step (2): computing weighted variance sum S of kth frame_kAs shown in the following formula:

S_k:＝0.5V_k+0.3V_k-1+0.2V_k-2

wherein the content of the first and second substances,

and (3): calculating S_kIs squared and normalized to [0,255 ]]Interval, get

Setting a threshold value T, dividing the value according to T

And obtaining a binary mask from the pixels in the image.

4. The automatic detection and tracking method of claim 1 wherein the target mask of the previous M frames is found by using a weighted sliding variance method, and the optimal mask is the mask with the largest proportion of foreground regions.

5. The automatic detection and tracking method of claim 1 wherein the adaptive projection is performed using an optimal mask, comprising the steps of:

step (1): performing horizontal and vertical projection on the mask, and counting the sum of horizontal and vertical pixels, which is D_h,D_v；

Step (2): calculating the center position (P) of the optimal target_x,P_y)：

And (3): setting the threshold Q to 0.97, avoiding the influence of noise, and calculating the optimum scale (R)_w,R_h) The pseudo code is as follows:

6. the automatic detection and tracking method of claim 1 wherein the effective tracking is performed using a centroid method comprising the steps of:

step (1): image I of the best mask obtained by weighted sliding variance method, and ROI (x, y, ROI) of the target obtained by adaptive projection method_W,roi_H) Calculating the centroid position of the current frame

Initializing a tracker;

step (2): inputting the next frame image, and converting it into grayscale image I_next；

And (3): the inter-frame difference D within the ROI is calculated using the following equation:

D(i-x,j-y):＝|I_next(i,j)-I(i,j)|,x≤i≤x+roi_W,y≤j≤y+roi_H

and (4): calculating a segmentation threshold Th by adopting an OTSU method, and binarizing D by using the Th;

and (5): finding the centroid C by_x,C_y：

And (6): the coordinates (x, y) of the ROI are updated with the following formula:

and (7): and (5) returning the updated ROI and repeating the step (2) until the video reading is finished.

Images