Intelligent director and picture synthesis method of panoramic live broadcast system

Abstract

The invention relates to the technical field of director and picture synthesis methods, in particular to an intelligent director and picture synthesis method of a panoramic live broadcast system, which comprises the following steps: acquiring panoramic video data shot by a panoramic camera; obtaining user feedback data; based on the panoramic video data, generating a shot selection strategy by using a deep reinforcement learning algorithm; detecting a live broadcast key event in real time by using a target detection and tracking algorithm based on the panoramic video data; generating a picture shooting plan sequence according to the shot selection strategy and the live broadcast key event; performing intelligent editing and synthesis on the panoramic video data based on the picture shooting plan sequence; and outputting the synthesized live video. And updating parameters of the deep reinforcement learning algorithm based on the user feedback data, thereby realizing full automation of the program directing process, greatly improving the efficiency, and enabling the system to continuously learn from live broadcast experience through the deep reinforcement learning algorithm and gradually improve the decision-making ability of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of directing and picture synthesis methods, and more specifically, to an intelligent directing and picture synthesis method for a panoramic live broadcast system. Background Art

[0002] With the rapid development of live broadcast technology, panoramic live broadcast system has gradually become an emerging video dissemination method. Compared with traditional live broadcast, panoramic live broadcast can provide viewers with a more immersive viewing experience, making the audience feel as if they were at the scene. However, the panoramic live broadcast system still faces many challenges in practical application, the most prominent of which is how to achieve intelligent directing and picture synthesis.

[0003] Traditional panoramic live broadcast systems usually rely on manual directing, which is not only time-consuming and labor-intensive, but also difficult to capture all the wonderful moments in time. Limited by personal experience and reaction speed, manual directing often finds it difficult to make the best shot selection in complex and ever-changing live broadcast scenes. In addition, manual directing also finds it difficult to analyze and respond to audience preferences in real time, resulting in the live broadcast content possibly failing to meet audience expectations.

[0004] In recent years, some automated directing systems have begun to emerge in an attempt to address the limitations of manual directing. These systems usually use preset rules or simple image analysis techniques to select shots and synthesize images. However, this approach still has many shortcomings. First, preset rules are difficult to adapt to complex and changing live broadcast scenes, and are prone to mechanical and rigid shot switching. Second, simple image analysis technology is difficult to accurately capture key events and character emotions in live broadcasts, resulting in important images being overlooked. Finally, these systems lack learning and optimization capabilities, making it difficult to continuously improve their performance based on audience feedback.

[0005] Therefore, the existing panoramic live broadcast system, whether it relies on manual directing or adopts simple automation solutions, is difficult to meet the requirements of efficiency, intelligence and personalization at the same time. This urgently requires a panoramic live broadcast system that can intelligently direct and synthesize images to improve the live broadcast quality and audience experience. Summary of the invention

[0006] The intelligent broadcast directing and picture synthesis method of the panoramic live broadcast system of the present invention is proposed precisely to solve the above technical problems. The method aims to realize a fully automatic and intelligent broadcast directing process, accurately capture key events in the live broadcast, analyze audience preferences in real time, and continuously optimize its performance based on feedback. By innovatively combining deep reinforcement learning and computer vision technology, the method of the present invention can overcome the limitations of the existing technology and bring a qualitative leap to the panoramic live broadcast system.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] The intelligent broadcasting and picture synthesis method of the panoramic live broadcast system includes:

[0009] The acquisition steps include:

[0010] Obtain panoramic video data captured by a panoramic camera;

[0011] Obtain user feedback data;

[0012] Processing steps include:

[0013] Based on the panoramic video data, a shot selection strategy is generated using a deep reinforcement learning algorithm;

[0014] Based on the panoramic video data, using target detection and tracking algorithms to detect key live events in real time;

[0015] Generate a picture shooting plan sequence according to the lens selection strategy and the live broadcast key events;

[0016] Based on the picture shooting plan sequence, intelligently editing and synthesizing the panoramic video data;

[0017] Output steps include:

[0018] Output the synthesized live video;

[0019] Based on the user feedback data, update the parameters of the deep reinforcement learning algorithm.

[0020] Preferably, the generating the shot selection strategy by using the deep reinforcement learning algorithm specifically includes:

[0021] Construct a Q-learning model with audience image acceptance as the goal;

[0022] Based on the Q learning model, reinforcement learning is performed on the lens selection strategy of the camera;

[0023] According to the user feedback data, the parameters of the Q learning model are updated in real time.

[0024] Preferably, the real-time detection of key live events using a target detection and tracking algorithm specifically includes:

[0025] Performing target detection on the panoramic video data using the Faster-RCNN method;

[0026] Use the MDNet method to track the detected target;

[0027] Based on the target detection and tracking results, key live events are identified.

[0028] As a preference, it also includes:

[0029] Generate a human emotion recognition model using an algorithm that combines deep object detection with human posture estimation;

[0030] Based on the character emotion recognition model, detecting the emotional response of the character in the live broadcast;

[0031] The emotional response result is associated with the shot selection strategy as a reference for updating the deep reinforcement learning algorithm.

[0032] Preferably, generating a picture shooting plan sequence specifically includes:

[0033] Build the Actor network of the deep reinforcement learning model and input the state value of each shot in the current frame;

[0034] Based on the Actor network, a directing strategy for directing actions performed by each shot is calculated;

[0035] Build the Critic network of the deep reinforcement learning model and input the current state and current action value;

[0036] Based on the Critic network, calculate the current reward value generated by the current state;

[0037] Use the Monte Carlo method to predict the current action value;

[0038] A picture shooting plan sequence is generated according to the director strategy, the current reward value and the current action value.

[0039] Preferably, the intelligent editing and synthesis of the panoramic video data specifically includes:

[0040] Switching each shot based on the picture shooting plan sequence;

[0041] Switch each shot according to the preset shot switching rules;

[0042] The switched lens images are synthesized to generate the final live video.

[0043] As a preference, it also includes:

[0044] The convolutional neural network model is used to combine the face recognition results and emotion recognition results to judge the audience's acceptance of the picture;

[0045] Based on the audience acceptance, the shot selection strategy is adjusted.

[0046] Preferably, the lens selection strategy further includes:

[0047] Calculate the optimal shot for the next frame based on the prediction of the character's motion trajectory;

[0048] A step-by-step control method is used to accurately control the movement of the camera lens.

[0049] Preferably, the lens selection strategy considers the following multiple states:

[0050] The character’s emotional state;

[0051] Character movement status;

[0052] The current state of the lens;

[0053] Picture attractiveness state.

[0054] As a preference, an adaptive update mechanism is also included:

[0055] According to the user feedback data, adjusting the lens selection strategy in real time;

[0056] According to the new shooting results, the deep reinforcement learning model is updated;

[0057] According to the scene changes, the parameters of the target detection and tracking algorithm are dynamically adjusted.

[0058] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0059] First, it fully automates the broadcasting process, greatly improving efficiency and maintaining high-quality video output even in complex and ever-changing live broadcast scenarios. For example, in a live broadcast of a large-scale music festival, the system can automatically capture the wonderful performances on each stage, seamlessly switch between different perspectives, and present the best viewing experience for the audience.

[0060] Secondly, the method of the present invention has powerful learning and optimization capabilities. Through the deep reinforcement learning algorithm, the system can continuously learn from the live broadcast experience and gradually improve its decision-making ability. This adaptability enables the system to quickly adapt to different types of live broadcast content, whether it is a sports event, a concert or a press conference, it can perform well.

[0061] Furthermore, this method achieves a high degree of personalization. By analyzing audience feedback and emotions in real time, the system can dynamically adjust its shot selection strategy to meet the preferences of different audience groups. This personalized viewing experience greatly improves audience engagement and satisfaction, which is conducive to improving the commercial value of live broadcasts.

[0062] In addition, the method of the present invention also significantly improves the picture quality. Through advanced target detection and tracking algorithms, the system can accurately capture key figures and events in the live broadcast, ensuring that important scenes are not missed. At the same time, the picture synthesis technology based on deep learning can achieve smooth and natural shot transitions, avoiding the common stiff switching problems of traditional automation systems.

[0063] Finally, the method of the present invention has excellent scalability and adaptability. It is not only applicable to various types of panoramic live broadcasts, but can also be easily integrated into existing live broadcast systems, bringing revolutionary changes to the entire live broadcast industry.

[0064] In summary, the intelligent directing and picture synthesis method of the panoramic live broadcast system of the present invention innovatively solves many problems faced by the existing technology, achieves a perfect combination of efficiency, intelligence and personalization, opens up a new path for the development of panoramic live broadcast technology, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The figure is an overall flow chart of the method of the present invention.

[0066] Figure 2 A logical block diagram for generating a shot selection strategy for deep reinforcement learning of the present invention.

[0067] Figure 3 The following is a logic block diagram of the target detection and tracking algorithm of the present invention for detecting key events.

[0068] Figure 4 A logic block diagram generated for the picture shooting plan sequence of the present invention.

[0069] Figure 5 It is a logic block diagram of the intelligent editing and synthesis of the present invention.

[0070] Figure 6 It is a logic block diagram of the audience acceptance judgment mechanism of the present invention.

[0071] Figure 7 It is a logic block diagram of the adaptive update mechanism of the present invention. DETAILED DESCRIPTION

[0072] like Figure 1-7 As shown, the present invention provides an intelligent broadcasting and picture synthesis method for a panoramic live broadcast system. The method innovatively combines deep reinforcement learning and computer vision technology to achieve intelligent and automated panoramic live broadcast. The technical solution of the present invention will be described in detail below.

[0073] First, the method of the present invention includes an acquisition step. In this step, the system acquires panoramic video data captured by a panoramic camera. These data are usually high-resolution, 360-degree video streams without blind spots. At the same time, the system also acquires user feedback data. User feedback data may include information such as viewing time, number of likes, and comment content, which reflect the audience's preference for the live content.

[0074] Next comes the processing step. In this step, the method of the present invention first generates a shot selection strategy based on panoramic video data using a deep reinforcement learning algorithm. The core of the deep reinforcement learning algorithm is that it can gradually optimize the decision-making process through continuous trial and learning. In the present invention, the algorithm learns how to choose the shots that are most attractive to the audience.

[0075] At the same time, the method of the present invention also uses the target detection and tracking algorithm based on the panoramic video data to detect key events in the live broadcast in real time. This step is crucial for capturing the wonderful moments in the live broadcast. For example, in the live broadcast of sports events, the algorithm can accurately identify key events such as goals and scores.

[0076] Then, according to the generated shot selection strategy and the detected live broadcast key events, the method of the present invention generates a picture shooting plan sequence, which determines how to switch and combine different shots during the live broadcast process to present the most attractive pictures.

[0077] Based on the generated picture shooting plan sequence, the method of the present invention performs intelligent editing and synthesis of the panoramic video data. This step converts the original panoramic video into carefully arranged live content.

[0078] In the output step, the system outputs the synthesized live video. This is the live content that the audience ultimately sees. At the same time, the system also updates the parameters of the deep reinforcement learning algorithm based on user feedback data. This feedback mechanism enables the system to continuously learn and improve, providing an increasingly better viewing experience.

[0079] Furthermore, when the method of the present invention generates a shot selection strategy using a deep reinforcement learning algorithm, the method specifically includes the following steps: First, a Q learning model with audience acceptance of the picture as the goal is constructed. Q learning is a classic algorithm in reinforcement learning, which optimizes the decision-making process by learning the action value function. In the present invention, the goal of the Q learning model is to maximize the audience's acceptance of the picture.

[0080] Next, based on the Q-learning model, the camera's lens selection strategy is reinforced. In this process, the system will continuously try different lens selection strategies and adjust the strategy based on the audience's response. For example, if a certain lens selection leads to a significant increase in the audience's viewing time, the system will tend to choose such a lens in similar situations.

[0081] Finally, the parameters of the Q-learning model are updated in real time based on user feedback data. This real-time update mechanism enables the system to quickly adapt to changes in audience preferences. For example, if the system detects that the audience's preference for a certain type of shot has changed, it will quickly adjust the corresponding parameters to provide live content that better suits the audience's taste.

[0082] The method of the present invention uses advanced computer vision technology when using target detection and tracking algorithms to detect key events in live broadcasts in real time. Specifically, the Faster-RCNN method is first used to detect targets in panoramic video data. Faster-RCNN is an efficient target detection algorithm that can quickly and accurately identify and locate various objects in real-time video streams.

[0083] After the target is detected, the method of the present invention uses the MDNet method to track the detected target. MDNet is a target tracking algorithm based on deep learning, which can stably track the target object in complex backgrounds. This is very important for keeping an eye on important people or objects in the live broadcast.

[0084] Finally, based on the target detection and tracking results, the system identifies key events in the live broadcast. For example, in a live concert, the system may identify the singer walking onto the stage as a key event, thereby triggering the corresponding camera switch.

[0085] The method of the present invention also includes an innovative step, which is to generate a character emotion recognition model using an algorithm that combines deep object detection with human posture estimation. This model can infer the emotional state of characters from their facial expressions and body language. For example, in a live sports game, this model can identify the excitement or frustration of athletes, thus providing an important reference for shot selection.

[0086] Based on the generated character emotion recognition model, the system is able to detect the emotional reactions of characters in live broadcasts. These emotional reaction results are associated with the shot selection strategy and serve as an important reference for updating the deep reinforcement learning algorithm. For example, if the system detects that a particular shot often captures the audience's excitement, it may increase the probability of selecting that shot.

[0087] Through the above detailed description, it can be seen how the method of the present invention cleverly combines a variety of advanced technologies to achieve intelligent directing and picture synthesis of a panoramic live broadcast system. This method not only greatly improves the quality and attractiveness of the live broadcast, but also can continuously optimize itself according to audience feedback to provide viewers with an increasingly better viewing experience. In a preferred embodiment of the present invention, the generation process of the picture shooting plan sequence involves the subtle application of a deep reinforcement learning model. Specifically, the process first constructs an Actor network of a deep reinforcement learning model. The network receives as input the state value of each lens of the current frame, which may include parameters such as the position, angle, and focal length of the lens.

[0088] Based on the Actor network, this method calculates the directing strategy of the directing action performed by each shot. The directing strategy here can be understood as a probability distribution, which represents the tendency to choose various possible actions. For example, in a live broadcast of a football game, when the ball is suddenly passed near the goal, the system may tend to choose a closer and more focused shot to capture the possible goal moment.

[0089] At the same time, the method of the present invention also constructs a Critic network of the deep reinforcement learning model. The role of this network is to evaluate the value of the current state and action. It receives the current state and the current action value as input and outputs an estimated value. In a live broadcast scenario, this value may represent the expected contribution of a certain shot selection to the audience's appeal.

[0090] Based on the Critic network, this method calculates the current reward value generated by the current state. This reward value reflects the immediate effect of the current shot selection. For example, if a shot selection leads to a significant increase in audience engagement, then this choice will receive a higher reward value.

[0091] Next, the method of the present invention uses the Monte Carlo method to predict the current action value. The Monte Carlo method is a numerical calculation method based on random sampling, which is used here to estimate the long-term cumulative return. This method can help the system consider not only the immediate return but also the long-term impact when making decisions.

[0092] Finally, based on the calculated directing strategy, current reward value, and current action value, this method generates a picture shooting plan sequence. This sequence determines how the system will switch and combine different shots in the following live broadcast process to present the most attractive picture.

[0093] In another embodiment of the present invention, the intelligent editing and synthesis process of panoramic video data is also worth describing in detail. First, based on the previously generated picture shooting plan sequence, the system switches each shot. This process requires precise time control and smooth transition effects to ensure that the audience's viewing experience is not affected by frequent shot switching.

[0094] When switching shots, this method will operate according to the preset shot switching rules. These rules may include the minimum switching interval, the maximum allowed switching speed, and the forced switching triggered by specific events. For example, in a live broadcast of a concert, the system may set a rule to automatically switch to a frontal close-up shot when the singer starts singing, and switch to a panoramic shot to show the entire stage effect during the interlude.

[0095] After the lens switching is completed, the method of the present invention synthesizes the switched lens images to finally generate a live video. This synthesis process may involve multiple aspects such as seamless splicing of images, coordinated unification of colors, and synchronous processing of audio to ensure that the live video outputted in the end has professional quality.

[0096] Preferably, the method of the present invention also includes a convolutional neural network-based audience acceptance judgment mechanism. The mechanism uses a convolutional neural network model and cleverly combines face recognition results and emotion recognition results to judge the audience's acceptance of the picture. This multimodal analysis method can capture the audience's response more comprehensively and accurately.

[0097] Specifically, facial recognition technology can help the system locate and track the faces of viewers, while emotion recognition technology can infer the emotional state of viewers from their facial expressions. For example, in a live broadcast of a sports game, if the system detects that most viewers show excited expressions at a certain exciting moment, then the picture will be judged to have high acceptance.

[0098] Based on the judgment results of the audience's acceptance, this method can adjust the shot selection strategy in time. If a certain type of shot or a specific shot transition frequently causes a positive response from the audience, the system will increase the probability of selecting this type of shot or using this type of transition. Conversely, if a certain shot selection often causes a decrease in audience interest, the system will reduce its frequency of use accordingly.

[0099] In this way, the method of the present invention realizes a closed-loop optimization process, which can continuously optimize the live broadcast content according to the real-time reaction of the audience, and provide a viewing experience that is more and more in line with the audience's taste. This intelligent and personalized live broadcast method will undoubtedly greatly enhance the attractiveness and value of panoramic live broadcast. In another preferred embodiment of the present invention, the process of formulating the lens selection strategy also includes an innovative mechanism based on the prediction of the character's motion trajectory. This mechanism can predict the position where the character may appear in the next frame, so as to calculate the optimal lens selection in advance. For example, in a live broadcast of a basketball game, if the system predicts that a player is about to catch the ball and may shoot, it will select a lens angle in advance that is more suitable for capturing this wonderful moment.

[0100] Specifically, this method uses advanced motion prediction algorithms, such as Kalman filters or long short-term memory networks (LSTM), to analyze the historical movement trajectory of a person and predict their future position. This prediction not only takes into account the current position and speed of the person, but also the particularity of the scene. For example, in a live concert, the system will predict the possible movement of the singer to ensure that the best performance can always be captured.

[0101] In addition, the method of the present invention also uses a step-by-step control method to accurately control the movement of the camera lens. This control method can achieve a smooth transition of the lens and avoid the discomfort of the audience caused by sudden large movements. In practical applications, the system will calculate the angle and distance that the lens needs to move based on the predicted position of the person, and then decompose this large movement into multiple small steps, each of which is very subtle, so as to achieve seemingly continuous and smooth lens movement.

[0102] Preferably, the method of the present invention comprehensively considers a variety of state information when formulating the shot selection strategy to achieve a more intelligent and comprehensive decision. These states include the emotional state of the character, the movement state of the character, the current state of the shot, and the attractiveness state of the picture.

[0103] The judgment of the emotional state of a person is based on the emotion recognition technology mentioned above. For example, in a live broadcast of an award ceremony, if the system detects that the winner is showing excitement, it may choose a closer close-up shot to capture this touching moment.

[0104] The character's motion status is based on the results of the target detection and tracking algorithm. This status information is particularly important in live sports events. For example, when the system detects that a player is making a fast break, it will choose a wide-angle lens that can fully show the action.

[0105] The current state of the lens includes parameters such as the lens position, angle, focal length, etc. When the system makes a new lens selection, it will take the current lens state into consideration to ensure that the transition between the old and new lenses is smooth and natural.

[0106] The attractiveness of a picture is based on the audience acceptance judgment mechanism described earlier. If a certain type of picture continues to receive high acceptance, the system will increase the tendency to select similar pictures.

[0107] By comprehensively considering these status information, the method of the present invention can make more intelligent and personalized shot selection decisions, thereby providing more attractive live broadcast content.

[0108] In another embodiment of the present invention, an adaptive update mechanism is also included, which enables the entire system to continuously optimize itself and adapt to different live broadcast scenarios and audience preferences.

[0109] First, the method adjusts the shot selection strategy in real time based on user feedback data. The user feedback data here may include indicators such as viewing time, number of likes, and density of bullet comments. For example, if the system finds that a certain type of shot transition often triggers active interaction from the audience, it will increase the frequency of use of such transitions.

[0110] Second, based on new shooting results, this method continuously updates the deep reinforcement learning model. This means that the system can learn from each live broadcast and gradually improve its decision-making ability. For example, if an innovative shot selection in a live broadcast receives unexpected praise, the system will incorporate this choice into its decision-making model and increase the probability of using this choice in similar situations in the future.

[0111] Finally, the method of the present invention can also dynamically adjust the parameters of the target detection and tracking algorithm according to scene changes. This adjustment ensures that the system can maintain efficient and accurate performance in different live broadcast environments. For example, in outdoor live broadcasts where light conditions vary greatly, the system may automatically adjust the threshold parameters of the target detection algorithm to adapt to the lighting conditions at different times.

[0112] Through this multi-level, multi-angle adaptive update mechanism, the method of the present invention can continuously optimize its performance and provide viewers with an increasingly high-quality panoramic live broadcast experience. Whether in complex and ever-changing sports events or exciting concerts, this intelligent broadcast director system can accurately capture every moment worth paying attention to and present professional live broadcast content to the audience.

[0113] The above description is only a preferred specific implementation manner of the present invention; however, the protection scope of the present invention is not limited thereto; any person familiar with the art who, within the scope disclosed by the present invention, makes equivalent replacements or changes based on the scheme and improved concepts of the present invention shall be covered within the protection scope of the present invention.

Claims

1. An intelligent broadcasting and picture synthesis method for a panoramic live broadcast system, characterized in that: include: The acquisition steps include: Obtain panoramic video data captured by a panoramic camera; Obtain user feedback data; Processing steps include: Based on the panoramic video data, a shot selection strategy is generated using a deep reinforcement learning algorithm; Based on the panoramic video data, using target detection and tracking algorithms to detect key live events in real time; Generate a picture shooting plan sequence according to the lens selection strategy and the live broadcast key events; Based on the picture shooting plan sequence, intelligently editing and synthesizing the panoramic video data; Output steps include: Output the synthesized live video; Based on the user feedback data, update the parameters of the deep reinforcement learning algorithm.

2. The method according to claim 1, characterized in that: The method of generating a shot selection strategy using a deep reinforcement learning algorithm specifically includes: Construct a Q-learning model with audience image acceptance as the goal; Based on the Q learning model, reinforcement learning is performed on the lens selection strategy of the camera; According to the user feedback data, the parameters of the Q learning model are updated in real time.

3. The method according to claim 1, characterized in that The real-time detection of key live events using the target detection and tracking algorithm specifically includes: Performing target detection on the panoramic video data using the Faster-RCNN method; Use the MDNet method to track the detected target; Based on the target detection and tracking results, key live events are identified.

4. The method according to claim 3, characterized in that Also includes: Generate a human emotion recognition model using an algorithm that combines deep object detection with human posture estimation; Based on the character emotion recognition model, detecting the emotional response of the character in the live broadcast; The emotional response result is associated with the shot selection strategy as a reference for updating the deep reinforcement learning algorithm.

5. The method according to claim 1, characterized in that: The generating picture shooting plan sequence specifically comprises: Build the Actor network of the deep reinforcement learning model and input the state value of each shot in the current frame; Based on the Actor network, a directing strategy for directing actions performed by each shot is calculated; Build the Critic network of the deep reinforcement learning model and input the current state and current action value; Based on the Critic network, calculate the current reward value generated by the current state; Use the Monte Carlo method to predict the current action value; A picture shooting plan sequence is generated according to the director strategy, the current reward value and the current action value.

6. The method according to claim 1, characterized in that The intelligent editing and synthesis of panoramic video data specifically includes: Switching each shot based on the picture shooting plan sequence; Switch each shot according to the preset shot switching rules; The switched lens images are synthesized to generate the final live video.

7. The method according to claim 1, characterized in that Also includes: The convolutional neural network model is used to combine the face recognition results and emotion recognition results to judge the audience's acceptance of the picture; Based on the audience acceptance, the shot selection strategy is adjusted.

8. The method according to claim 1, characterized in that The lens selection strategy also includes: Calculate the optimal shot for the next frame based on the prediction of the character's motion trajectory; A step-by-step control method is used to accurately control the movement of the camera lens.

9. The method according to claim 1, characterized in that: The lens selection strategy considers the following multiple states: The character’s emotional state; Character movement status; The current state of the lens; Picture attractiveness state.

10. The method according to claim 1, characterized in that Also includes an adaptive update mechanism: According to the user feedback data, adjusting the lens selection strategy in real time; According to the new shooting results, the deep reinforcement learning model is updated; According to the scene changes, the parameters of the target detection and tracking algorithm are dynamically adjusted.

Images