CN105844664A - Monitoring video vehicle detection tracking method based on improved TLD - Google Patents

Abstract

The invention discloses a monitoring video vehicle detection tracking method based on an improved TLD, which is high in accuracy and good in robustness. The monitoring video vehicle detection tracking method comprises steps of adopting a sub-block Cam Shift tracer based on vehicle color characteristics to replace an L-K light flow point tracer, realizing description of a tracked object through a vehicle area color histogram obtained by the Cam Shift, realizing activity amount estimation on the tracking object moving between a prior frame and a later frame through measuring the similarity of the color histograms in the capture area, further combining with a random forest detector to obtain a rough position of a vehicle object, and observing a detector and positioning the tracker in real time through P-N study so as to realize effective vehicle detection tracking.

Description

Surveillance video vehicle detection and tracking method based on improved TLD

technical field

The invention relates to the field of intelligent traffic video processing, in particular to a monitoring video vehicle detection and tracking method based on improved TLD with high accuracy and good robustness.

Background technique

The frame-by-frame tracking of the target by the L-K optical flow method is performed under the assumption that no target disappears or is completely occluded, which is also called a short-term tracker. This type of tracker usually lacks direct processing after tracking errors occur, and it is difficult to achieve good results in long-term target tracking. At present, people's research on this kind of short-term tracking method mainly focuses on improving the tracking accuracy and speed, and prolonging the tracking time. However, when the tracking accuracy is not ideal, the accumulation of tracking errors and drift cannot be effectively avoided. In recent years, a new single-target long-term tracking algorithm TLD (Tracking Learning - Detection) has emerged, which combines tracking algorithms and detection algorithms to overcome the problems of deformation and partial occlusion of the target during the tracking process; At the same time, the algorithm introduces an online learning mechanism, which inputs the results obtained by the tracker and detector to the learning module, and feeds the learned model back to the tracking and detection module, so that the detection and tracking of the target are more stable and effective . However, since the tracking module uses the L-K optical flow method based on the assumption of motion consistency and coherence, the algorithm has a better prediction result for the target's motion between consecutive frames when the cycle is short, the motion between frames is limited, and the naked eye is visible. , but for those large and fast moving targets, its prediction and tracking performance is not ideal yet.

Contents of the invention

The purpose of the present invention is to solve the above-mentioned technical problems existing in the prior art, and to provide a monitoring video vehicle detection and tracking method based on the improved TLD with high accuracy and good robustness.

The technical solution of the present invention is: a kind of surveillance video vehicle detection and tracking method based on improved TLD, it is characterized in that carrying out according to the following steps:

Step 1. Input the first frame of video image, manually mark the target to be tracked, make ;

Step 2. Initialize the random forest classifier and Cam Shift tracker;

Step 3. Order , load the first Frame the video image, and use the random forest classifier to detect the target, use the CamShift tracker to track the target and get the adjusted scale of the target frame;

Step 4. Fuse the detection results of the random forest classifier with the tracking results of the Cam Shift tracker;

Step 5. Use the P-N learning strategy to update the random forest classifier to obtain the position of the target;

Step 6. If the video has reached the last frame, the algorithm ends; otherwise, go to Step 3.

The Step 4 is as follows:

Step 4.1 If both the random forest classifier and the Cam Shift tracker have bounding boxes as output, but the random forest classifier has multiple similar positions determined, while the Cam Shift tracker only finds one target position, at this time the spatial overlap Perform clustering and segmentation on several detection results;

Step 4.2 If the Cam Shift tracker has no bounding box output, but the random forest classifier has a bounding box output, then multiple detection results are segmented by clustering with spatial overlap, and the first cluster segmentation result is used as the fusion result;

Step 4.3 If a clustering decision result with a large correlation value appears, but the decision result is far from the result of the Cam Shift tracker, use the decision result as the fusion result, and then re-initialize the Cam Shift tracker and discard the original The sample set that is considered correct;

Step 4.4 If the Cam Shift tracker has a bounding box output, but the Random Forest classifier has no bounding box output, then use the result output by the Cam Shift tracker as the fusion result;

Step 4.5 If neither the Random Forest classifier nor the Cam Shift tracker outputs a bounding box, the target is considered to have disappeared.

The present invention adopts the segmented Cam Shift tracker based on vehicle color features to replace the point tracker of L-K optical flow, realizes the description of the tracking target through the color histogram of the vehicle area obtained by Cam Shift, and then captures the similarity of the color histogram of the area Realize the prediction of the motion of the tracking target between the two frames before and after the tracking target; further combine the random forest detector to obtain the rough position of the vehicle target, and observe the detector and position the tracker in real time through P-N learning, so as to achieve effective vehicle detection and tracking. Compared with the prior art, the present invention improves the tracking accuracy and robustness of the moving vehicle under large and rapid changes in the long-time tracking process.

Description of drawings

Fig. 1 is a flowchart of an embodiment of the present invention.

Figure 2 is a comparison of the detection and tracking results in the urban surveillance scene at an oblique angle.

Figure 3 is a comparison of detection and tracking results in high-altitude city surveillance video scenarios.

Figure 4 is a comparison chart of detection and tracking results in a rainy urban road surveillance video scene.

Figure 5 is a comparison of the detection and tracking results in the surveillance video scene of the oblique expressway.

detailed description

As shown in Figure 1: The steps of the monitoring video vehicle detection and tracking algorithm based on the improved TLD are as follows:

Step 1. Input the first frame of video image, and manually mark the target to be tracked manually, so that ;

Step 2. Initialize the random forest classifier and Cam Shift tracker;

Step 3. Order , load the first Frame the video image, use the random forest classifier to detect the target, and use the CamShift tracker to track the target, so as to obtain the adjusted scale of the target frame;

Step 4. Fuse the detection results of the random forest classifier with the tracking results of the Cam Shift tracker;

Step 5. Use the P-N learning strategy to update the random forest classifier to obtain the position of the target;

Step 6. If the video has reached the last frame, the algorithm ends; otherwise, go to Step 3.

Described Step 4 comprises the following steps:

Step 4.1 If both the random forest classifier and the Cam Shift tracker have bounding boxes as output, but the random forest classifier has multiple similar positions determined, while the Cam Shift tracker only finds one target position, at this time the spatial overlap Perform clustering and segmentation on several detection results;

Step 4.2 If the Cam Shift tracker has no bounding box output, but the random forest classifier has a bounding box output, then multiple detection results are segmented by clustering with spatial overlap, and the first cluster segmentation result is used as the fusion result;

Step 4.3 If a clustering decision result with a large correlation value appears, but the decision result is far from the result of the Cam Shift tracker, use the decision result as the fusion result, and then re-initialize the Cam Shift tracker and discard the original The sample set that is considered correct;

Step 4.4 If the Cam Shift tracker has a bounding box output, but the Random Forest classifier has no bounding box output, then use the result output by the Cam Shift tracker as the fusion result;

Step 4.5 If neither the Random Forest classifier nor the Cam Shift tracker outputs a bounding box, the target is considered to have disappeared.

The comparison of the detection and tracking results in the urban monitoring scene at an inclination angle according to the embodiment of the present invention is shown in FIG. 2 .

The comparison of the detection and tracking results in the high-altitude city surveillance video scene of the embodiment of the present invention is shown in FIG. 3 .

The comparison of the detection and tracking results in the monitoring video scene of urban roads in rainy days according to the embodiment of the present invention is shown in FIG. 4 .

The comparison of the detection and tracking results in the monitoring video scene of the oblique expressway in the embodiment of the present invention is shown in FIG. 5 .

The comparison of program running time in the above-mentioned different scenarios according to the embodiment of the present invention is shown in Table 1:

Table 1

The tracking quality comparison of the embodiments of the present invention in the above-mentioned different scenarios is shown in Table 2:

Table 2

Claims (2)

1. a monitor video vehicle detecting and tracking method based on improvement TLD, it is characterised in that follow the steps below:

Step 1. inputs the 1st frame video image, and hand labeled goes out target to be tracked, order；

Step 2. initializes random forest grader and Cam Shift tracker；

Step 3. makes, it is loaded into theFrame video image, and utilize random forest detection of classifier target, utilize Cam Shift tracker is followed the tracks of target and obtains the adjustment yardstick of target frame；

The tracking result of the testing result of random forest grader with Cam Shift tracker is blended by Step 4.；

Step 5. utilizes P-N learning strategy to update random forest grader, it is thus achieved that the position of target；

If Step 6. video has arrived at last frame, then algorithm terminates；Otherwise, Step 3 is proceeded to.

The most according to claim 1 based on the monitor video vehicle detecting and tracking method improving TLD, it is characterised in that described Step 4 is as follows:

If Step 4.1 random forest grader and Cam Shift tracker all bounding box are as output, but random forest Grader has multiple analogous location to be determined out, and Cam Shift tracker only finds a target location, now with space Degree of overlapping carries out cluster segmentation to some testing results；

If Step 4.2 Cam Shift tracker does not has bounding box to export, and the output of random forest grader bounding box, So multiple testing results are split with the cluster of space overlap degree, now use first cluster segmentation result as melting Close result；

If the bigger Clustering Decision-Making result of 4.3 1 correlations of Step occurs, but this result of decision and Cam Shift with Track device result differ farther out, then use this result of decision as fusion results, at the beginning of then Cam Shift tracker being re-started Beginningization also loses the sample set originally praised；

If Step 4.4 Cam Shift tracker bounding box exports, and random forest grader borderless easel output, So use the result of Cam Shift tracker output as fusion results；

If Step 4.5 random forest grader and the output of Cam Shift tracker equal borderless easel, then it is assumed that target disappears Lose.