Video monitoring holder control method

Abstract

The invention relates to the field of monitoring holders, in particular to a video monitoring holder control method, which comprises the following steps: collecting continuous frames of monitoring images, obtaining a multi-dimensional data set of a target person, training a face network according to a face information set, and similarly obtaining an upper body network, a lower body network and a foot network; normalizing the calculated face distinguishing coefficient, the calculated upper body distinguishing coefficient, the calculated lower body distinguishing coefficient and the calculated foot distinguishing coefficient to serve as weights; obtaining a face recognition probability, an upper body recognition probability, a lower body recognition probability and a foot recognition probability; and carrying out weighted summation on the face recognition probability, the upper body recognition probability, the lower body recognition probability and the foot recognition probability to obtain a comprehensive recognition probability so as to control the monitoring cloud platform to monitor the target person. According to the technical scheme of the invention, the precision of the control result of the video monitoring holder can be improved.

Description

Technical Field

[0001] The present invention relates to the field of monitoring pan-tilt platforms, and more particularly to a method for controlling a video monitoring pan-tilt platform. Background Art

[0002] Video surveillance systems are used in a variety of industries, including public security, traffic management, and commerce. Traditional surveillance systems typically rely on fixed camera angles, lacking flexibility and making it difficult to fully cover the monitored area. Modern video surveillance systems are increasingly becoming more intelligent and flexible, with pan / tilt control becoming a key technology for improving system performance and flexibility.

[0003] Traditional video surveillance PTZ control methods struggle to effectively track target individuals in crowded environments. This is primarily due to the fact that as a target person moves, their physical features may become unclear or obscured, preventing traditional target recognition methods from accurately capturing target information. Furthermore, existing methods often rely on a single physical feature for recognition. This single-feature approach can easily lead to confusion when faced with individuals with similar features, resulting in misidentification or missed recognition, reducing the accuracy of monitoring results and causing inaccurate control. Summary of the Invention

[0004] In order to solve the technical problem of inaccurate control results, the present invention provides a video surveillance pan-tilt control method. It includes: collecting continuous frames of surveillance images, taking any person as the target person, obtaining a multidimensional data set of the target person, the multidimensional data set including a face information set, an upper body information set, a lower body information set and a foot information set, training a face network based on the face information set, and similarly obtaining an upper body network, a lower body network and a foot network; selecting a frame of surveillance image with the most pixels of the target person, calculating the face distinction coefficient of the target person, and similarly obtaining an upper body distinction coefficient, a lower body distinction coefficient and a foot distinction coefficient according to the calculation method of the face distinction coefficient, and then summing the face distinction coefficient, the upper body distinction coefficient and the lower body distinction coefficient to obtain the face distinction coefficient. The discrimination coefficient and the foot discrimination coefficient are normalized as weights respectively; the real-time monitoring image is segmented to obtain the face and upper body search radius, upper and lower body search radius, and lower body and foot search radius of the target person to complete the part matching of the target person, and input them into the face network, upper body network, lower body network, and foot network respectively to obtain the face recognition probability, upper body recognition probability, lower body recognition probability, and foot recognition probability; the face recognition probability, upper body recognition probability, lower body recognition probability, and foot recognition probability are weighted and summed to obtain the comprehensive recognition probability, so as to control the monitoring of the target person by the monitoring pan-tilt station.

[0005] Preferably, the training of the face network based on the face information set includes: taking the face information set of any person in history as input and the true value of the face recognition probability of any person in history as a label to obtain a training set, calculating the loss value between the predicted value and the true value of the face recognition probability of the face network, the loss function is the cross entropy loss, and when the loss value is less than the preset loss value or reaches the preset number of training times, the training is stopped to complete the training of the face network.

[0006] Preferably, the facial distinction coefficient includes: taking persons other than the target person as control persons, obtaining the facial pixel matrices of the target person and the control persons respectively, calculating the similarity between the target person and the control persons, traversing to obtain the similarity between the target person and each control person, and taking the average accumulated value of all similarities as the facial distinction coefficient.

[0007] Preferably, the face distinction coefficient also includes: a preset pixel threshold, retaining surveillance images in which the pixels in the face area are not less than the preset pixel threshold, calculating the average cumulative value of the corresponding similarity in each surveillance image, and taking the average of the average cumulative values ​​of all surveillance images as the face distinction coefficient.

[0008] Preferably, the method of obtaining the face and upper body search radius, the upper and lower body search radius, and the lower body and feet search radius of the target person includes: selecting a frame of surveillance image with the most pixels of the target person as a standard image, obtaining the number of face pixels in the standard image, and calculating the Euclidean distance between the face center and the upper body center of the target person in the standard image; calculating the ratio of the number of face pixels in the real-time surveillance image to the number of face pixels in the standard image, and taking the product of the ratio and the Euclidean distance as the face and upper body search radius; and obtaining the upper and lower body search radius and the lower body and feet search radius in the same way according to the method for obtaining the face and upper body search radius.

[0009] Preferably, the part matching of the target person is completed by: obtaining the face and upper body search radius, the upper and lower body search radius and the lower body and feet search radius to obtain the face and upper body search circle, the upper and lower body search circle and the lower body and feet search circle; for any face and upper body search circle, obtaining the pixel point intersection of each upper and lower body search circle and any face and upper body search circle, and matching the upper and lower body search circle with the largest intersection with any face and upper body search circle; similarly, obtaining the pixel point intersection of each lower body and feet search circle and any face and upper body search circle, and matching the lower body and feet search circle with the largest intersection with any face and upper body search circle to complete the part matching of the target person.

[0010] Preferably, the comprehensive recognition probability satisfies the relationship: , represents the comprehensive recognition probability, Indicates the location The weight of Indicates the location The recognition probability of Represents a face, Indicates upper body, Indicates lower body, Indicates the feet.

[0011] Beneficial effects of the present invention: The present invention effectively improves the ability to accurately monitor target personnel through the establishment of a multidimensional data set and the training of various part recognition networks. By independently identifying and weighting the target person's face, upper body, lower body, and feet, it ensures that the pan / tilt head can accurately focus and adjust the viewing angle during real-time monitoring, thereby optimizing the tracking effect. Specifically, by calculating the distinction coefficients and recognition probabilities of the face, upper body, lower body, and feet, dynamic monitoring of the target person is achieved, avoiding the problems of misidentification or deviation from the target in traditional methods. This helps to improve the response speed and recognition accuracy of the monitoring system in complex environments, can better adapt to the personnel monitoring needs in different scenarios, and enhances the intelligence and flexibility of the monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention provides a flow chart of a video surveillance PTZ control method. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0014] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] Reference Figure 1 A video surveillance PTZ control method includes steps S1 to S4, specifically as follows: The scenario of the present invention is that multiple cameras are controlled by a monitoring pan-tilt system to monitor the continuous behavior of any target person.

[0016] S1: Collect surveillance images of continuous frames, select any person as the target person, and obtain a multidimensional dataset of the target person. The multidimensional dataset includes a face information set, an upper body information set, a lower body information set, and a foot information set. The face network is trained based on the face information set, and the upper body network, lower body network, and foot network are obtained similarly.

[0017] In one embodiment, a camera at an entrance captures continuous frames of surveillance images. From each image, a multidimensional dataset of the target person is extracted, including facial information, upper body information, lower body information, and foot information. These datasets represent the target person's physical characteristics from different perspectives, providing a more comprehensive identity profile. Based on this, by analyzing the facial information sets in the continuous frames, a specialized face recognition network can be trained to identify and track the target person's facial features.

[0018] The specific training process is as follows: taking the facial information set of any person in history as input and the true value of the face recognition probability of any person in history as the label to obtain the training set, calculating the loss value between the predicted value and the true value of the face recognition probability of the face network, and the loss function is the cross entropy loss. When the loss value is less than the preset loss value or reaches the preset number of training times, the training is stopped and the training of the face network is completed.

[0019] For example, when the loss value is less than 0.001 or the number of training times reaches 200, the training is stopped.

[0020] Similarly, the upper body network, lower body network and foot network of the target person can be obtained.

[0021] S2: Select a surveillance image frame with the most pixels of the target person, calculate the face distinction coefficient of the target person, and obtain the upper body distinction coefficient, lower body distinction coefficient, and foot distinction coefficient according to the calculation method of the face distinction coefficient. Then normalize the face distinction coefficient, upper body distinction coefficient, lower body distinction coefficient, and foot distinction coefficient respectively as weights.

[0022] In one embodiment, based on the continuous frames of surveillance images obtained in step S1, the frame containing the most pixels of the target person is selected as the reference image. Then, the other persons in the image except the target person are regarded as the control persons, and the facial pixel matrices of the target person and the control persons are extracted respectively.

[0023] By calculating the facial similarity between the target person and each control person, a similarity value is obtained, which reflects the degree of similarity between the target person and the control person. Next, the similarity value of each control person is calculated for each control person. Finally, the average accumulated value of these similarities is used as the facial discrimination coefficient. This coefficient effectively represents the difference in facial features between the target person and other people, thereby helping to improve the accuracy and discrimination of the face recognition system.

[0024] The upper body distinction coefficient, lower body distinction coefficient and foot distinction coefficient are obtained in the same way as the calculation method of the face distinction coefficient, and the face distinction coefficient, upper body distinction coefficient, lower body distinction coefficient and foot distinction coefficient are normalized as weights respectively.

[0025] S3: Segment the real-time surveillance image to obtain the target person's face and upper body search radius, upper and lower body search radius, and lower body and foot search radius, complete the target person's part matching, and input the face network, upper body network, lower body network, and foot network respectively to obtain the face recognition probability, upper body recognition probability, lower body recognition probability, and foot recognition probability.

[0026] It should be noted that in order to achieve continuous monitoring of the target person by multiple cameras, it is necessary to first process the real-time surveillance images captured by any camera other than the entrance, identify features such as the face, upper body, lower body, and feet, and compare these features with the pre-stored features of the target person through matching. During the matching process, special consideration must be given to occlusion issues. In complex environments, some of the target person's physical features may be obscured by other objects or people, resulting in the inability to fully extract or accurately match the features. After the matching is completed, the matching results are input into the network, and the network determines whether the result belongs to the target person.

[0027] In one embodiment, the real-time monitoring image is segmented.

[0028] Select the surveillance image frame with the most pixels of the target person as the standard image, obtain the number of facial pixels in the standard image, and calculate the Euclidean distance between the face center and the upper body center of the target person in the standard image; calculate the ratio of the number of facial pixels in the real-time surveillance image to the number of facial pixels in the standard image, and use the product of the ratio and the Euclidean distance as the face and upper body search radius; similarly, obtain the upper and lower body search radius and the lower body and feet search radius based on the method of obtaining the face and upper body search radius.

[0029] When matching target person parts, the system first determines the areas of the face-upper body search circle, upper-lower body search circle, and lower-body-foot search circle, respectively, based on the search radii for the face, upper body, lower body, and feet. These search circles are used to locate and match the feature information of different body parts. For each given face-upper body search circle, the system further determines the pixel intersection of each upper-lower body search circle with the face-upper body search circle. The upper-lower body search circle with the largest intersection is then selected to match the face-upper body search circle. This ensures that the matched upper and lower body features have a high degree of spatial overlap with those of the target face, improving matching accuracy.

[0030] Similarly, this process is repeated for each lower body and foot search circle, obtaining the pixel intersection with each face and upper body search circle. The lower body and foot search circle with the largest intersection is then selected for matching. Finally, by sequentially matching the search circles for the face, upper body, lower body, and feet, the target person's body parts are matched. This process maximizes spatial overlap to ensure the most accurate matching results for different body parts, thereby improving the accuracy and robustness of the person recognition system, especially in dynamic scenes or when multiple people are occluded.

[0031] At this point, several matching results can be obtained in the real-time monitoring image. The face in any matching result is input into the face network, the upper body is input into the upper body network, the lower body is input into the lower body network, and the feet are input into the foot network, thereby obtaining the face recognition probability, upper body recognition probability, lower body recognition probability and foot recognition probability.

[0032] S4: A comprehensive recognition probability is obtained by weighted summing the face recognition probability, the upper body recognition probability, the lower body recognition probability, and the foot recognition probability to control the monitoring of the target person by the monitoring platform.

[0033] In one embodiment, the comprehensive recognition probability satisfies the relationship: , represents the comprehensive recognition probability, Indicates the location The weight of Indicates the location The recognition probability of Represents a face, Indicates upper body, Indicates lower body, Indicates the feet.

[0034] That is, the face recognition probability, upper body recognition probability, lower body recognition probability and foot recognition probability finally obtained in step S3.

[0035] At this point, it is possible to determine whether the target person exists in the real-time monitoring image, and then control the monitoring pan-tilt head to monitor the target person.

[0036] It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be based on the appended claims.

Claims

1. A video surveillance PTZ control method, characterized in that: include: Collect continuous frames of surveillance images, select any person as the target person, and obtain a multidimensional dataset of the target person. The multidimensional dataset includes a face information set, an upper body information set, a lower body information set, and a foot information set. The face network is trained based on the face information set, and the upper body network, lower body network, and foot network are obtained similarly. Select the surveillance image frame with the most pixels of the target person, calculate the face discrimination coefficient of the target person, and obtain the upper body discrimination coefficient, lower body discrimination coefficient, and foot discrimination coefficient according to the calculation method of the face discrimination coefficient. Then normalize the face discrimination coefficient, upper body discrimination coefficient, lower body discrimination coefficient, and foot discrimination coefficient respectively as weights; Segment the real-time surveillance image to obtain the target person's face, upper body, lower body, and foot search radius, complete the target person's part matching, and input the face network, upper body network, lower body network, and foot network respectively to obtain the face recognition probability, upper body recognition probability, lower body recognition probability, and foot recognition probability; The weighted sum of the face recognition probability, upper body recognition probability, lower body recognition probability and foot recognition probability is used to obtain the comprehensive recognition probability to control the monitoring of the target person by the monitoring pan-tilt head.

2. A video surveillance PTZ control method according to claim 1, characterized in that: The training of the face network according to the face information set includes: The facial information set of any person in history is used as input and the true value of the face recognition probability of any person in history is used as the label to obtain the training set. The loss value between the predicted value and the true value of the face recognition probability of the face network is calculated. The loss function is the cross entropy loss. When the loss value is less than the preset loss value or reaches the preset number of training times, the training is stopped and the training of the face network is completed.

3. A video surveillance PTZ control method according to claim 1, characterized in that: The face distinction coefficient includes: The persons other than the target person are taken as control persons, and the facial pixel matrices of the target person and the control persons are obtained respectively. The similarity between the target person and the control persons is calculated, and the similarity between the target person and each control person is obtained by traversing. The average accumulated value of all similarities is taken as the face distinction coefficient.

4. A video surveillance PTZ control method according to claim 3, characterized in that: The face distinction coefficient also includes: A pixel threshold is preset, and surveillance images with face area pixels not less than the preset pixel threshold are retained. The average cumulative value of the corresponding similarity in each surveillance image is calculated, and the mean of the average cumulative values ​​of all surveillance images is used as the face distinction coefficient.

5. A video surveillance PTZ control method according to claim 1, characterized in that: The method of obtaining the face and upper body search radius, the upper and lower body search radius, and the lower body and feet search radius of the target person includes: Select the surveillance image with the most pixels of the target person as the standard image, obtain the number of facial pixels in the standard image, and calculate the Euclidean distance between the face center and the upper body center of the target person in the standard image; Calculate the ratio of the number of face pixels in the real-time monitoring image to the number of face pixels in the standard image, and use the product of the ratio and the Euclidean distance as the face upper body search radius; The upper and lower body search radius and the lower body and foot search radius are obtained similarly based on the method of obtaining the upper and lower body search radius of the face.

6. A video surveillance PTZ control method according to claim 1, characterized in that: The part matching of the target person is completed as follows: The face upper body search radius, the upper body lower body search radius and the lower body foot search radius are used to obtain the face upper body search circle, the upper body lower body search circle and the lower body foot search circle; For any face and upper body search circle, obtain the pixel intersection of each upper and lower body search circle with any face and upper body search circle, and match the upper and lower body search circle with the largest intersection with any face and upper body search circle; Similarly, the pixel intersection of each lower body and foot search circle and any face and upper body search circle is obtained, and the lower body and foot search circle with the largest intersection is matched with any face and upper body search circle to complete the part matching of the target person.

7. A video surveillance PTZ control method according to claim 1, characterized in that: The comprehensive recognition probability satisfies the relationship: , represents the comprehensive recognition probability, Indicates the location The weight of Indicates the location The recognition probability of Represents a human face, Indicates upper body, Indicates lower body, Indicates the feet.

Images