A real-time high-speed target recognition and tracking method based on a localized artificial intelligence framework

Abstract

The application discloses a real-time high-speed target recognition and tracking scheme based on a domestic artificial intelligence framework, realizes rapid target detection and accurate tracking by combining target recognition and tracking technology, can automatically restart the recognition network and reposition when the system loses the target, supports manual selection of target tracking by the user, adopts a self-developed long-time tracking algorithm, is not limited by the target category, processes multiple videos in the front end, optimizes the use of computing resources, and improves the processing speed and efficiency by optimizing the hardware visual preprocessing and network inference part. The system provides a flexible output mode and is suitable for various application scenarios.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and artificial intelligence technology, and specifically to a real-time, high-speed target recognition and tracking solution based on a domestically produced artificial intelligence framework. Background Art

[0002] Artificial intelligence target recognition and tracking technology is one of the important research directions in the field of modern computer vision. With the rapid development of deep learning and big data technology, AI has made significant progress in image and video processing. Target recognition and tracking technology aims to automatically detect, identify and continuously track specific targets in videos or image sequences through computer vision algorithms. Traditional target recognition and tracking methods rely on manually designed feature extraction and classifiers. These methods often perform poorly when dealing with complex scenes and changing environments. In recent years, models such as convolutional neural networks and recurrent neural networks based on deep learning have demonstrated excellent performance in target recognition and tracking tasks. By training large-scale neural network models, AI systems can automatically learn features in images and achieve high-precision target detection and classification. In terms of target tracking, deep learning combines multi-target tracking and single-target tracking technologies, which can maintain continuous tracking of targets in complex backgrounds and occlusions. Common tracking algorithms include correlation filters, twin networks, and neural networks. Networks and tracking methods based on reinforcement learning. These algorithms ensure accurate tracking of targets in dynamic environments by continuously updating the target's location information. AI target recognition and tracking technology has broad application prospects in many fields. For example, in intelligent monitoring systems, it can detect and track suspicious people or objects in real time to improve public safety. In the field of autonomous driving, it can identify and track dynamic targets such as pedestrians and vehicles on the road to improve driving safety. In drone navigation and robot vision, it can achieve autonomous obstacle avoidance and target tracking, and enhance autonomy and intelligence. In summary, artificial intelligence target recognition and tracking technology is continuously promoting the development of computer vision technology by combining advanced deep learning algorithms and big data analysis, and is playing an increasingly important role in all walks of life.

[0003] However, existing target recognition and tracking system frameworks suffer from the following issues: They often fail to meet target detection speed requirements and lack sufficient tracking accuracy. Furthermore, when a target is lost, manual re-positioning is often required, resulting in a low level of automation and susceptibility to target type limitations. Therefore, a corresponding technical solution is needed to address these issues. Summary of the Invention

[0004] The purpose of the present invention is to provide a real-time, high-speed target recognition and tracking solution based on a domestically produced artificial intelligence framework, which solves the technical problems that the target detection speed cannot meet the requirements and the target tracking accuracy is poor. In addition, when the system target is lost, manual repositioning is usually required, the degree of automation is low, and it is easily restricted by the target category.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a real-time high-speed target recognition and tracking solution based on a domestic artificial intelligence framework, including two network frameworks of target recognition and target tracking. The working logic is to use the recognition network to trigger the tracking network. When the recognition network recognizes a trusted target, it quickly switches to the tracking network to lock the target and subsequently accurately track and predict the track. When the target is lost from the field of view, the tracking stops. At this time, the recognition network is reactivated to search for suspected targets in the field of view. The target in the field of view is manually selected, and the tracking network is activated to lock the target and subsequently accurately track and predict the track. When the target is lost from the field of view, the tracking stops to prevent mistracking. A self-developed long-term tracking algorithm is used. The long-term tracker will During the tracking process, deep feature extraction is performed on the target. When the target disappears and re-enters the field of view, it has a persistent and controllable re-capture capability. After the front-end multi-channel video image input, the target recognition and target tracking networks are time-sharing switched, mutually activated and scheduled, and the hardware visual pre-processing core VPC is called to complete the cutout, scaling, overlay, splicing, histogram statistics, color remapping, format conversion, Remap transformation, and filtering image processing operations. The pre-processed data is sent to the NPU memory buffer for subsequent reasoning. After the network reasoning is completed, it is output in two forms, namely the coordinates of the center point of the identified target and the coordinate sequence of the identification box range. The system superimposes the reasoning results into the video stream encoding output.

[0006] As a preferred embodiment of the present invention, the target recognition and target tracking networks are in a time-sharing serial working mode, and the resources and time of each working module network are independent. After the pre-processing is completed, they are sent to the NPU for model inference.

[0007] As a preferred embodiment of the present invention, the software execution process of the target identification and tracking solution is that the UART / ETH control thread continuously monitors the serial port and network port data, and is responsible for processing the sending and receiving tasks of the serial port or network port. When the UART / ETH control thread receives the instruction, it will automatically parse the instruction, convert the instruction information into the corresponding tracker working status and synchronize it to the Decoder module. The subsequent video processing related modules perform corresponding processing based on the status information. When the detector and tracker report the detection and tracking information to the UART / ETH control thread, the UART / ETH control thread will process the reported information according to the instruction protocol and send it to the external master immediately after processing.

[0008] As a preferred embodiment of the present invention, the target recognition and tracking algorithm first decodes the input video in a single frame. The decoded image is identified, tracked or directly output according to the main control instructions. The processed image is superimposed with display information according to the protocol requirements. The OSD module outputs the image and sends it to the encoder for image encoding and sends it to the external main control through the RTSP network video stream.

[0009] As a preferred embodiment of the present invention, the target recognition thread algorithm includes two threads: recognition pre-processing, target recognition, and recognition post-processing. The target recognition thread pre-processes the input current frame image to meet the requirements of the recognition network input, sends it to the recognition thread for model inference, and obtains the image output. The recognition post-processing thread is responsible for matching the target ID, determining the target category, and outputting the target recognition accuracy based on the network output and the timeline information. If the target is lost and re-enters the field of view within 3 seconds, the previous ID can be automatically matched. At the same time, it has the function of counting the number of targets of different categories.

[0010] As a preferred embodiment of the present invention, the tracking algorithm thread is responsible for processing the template extraction and calculating the search area after the tracking instruction is issued. After determining the tracking template, the template and the search area are sent to the tracking network for model reasoning, and the position of the tracking target and the confidence of the tracking target are output.

[0011] As a preferred embodiment of the present invention, the tracking post-processing thread function is to judge the current tracker status and dynamically update the tracking template. Its main logic is: if the tracking output confidence is greater than the set threshold, the tracking is considered normal; if the tracking output confidence is less than the set threshold, the target is considered lost or blocked.

[0012] As a preferred embodiment of the present invention, the specific states of the tracking algorithm execution logic are divided into the following three states: normal tracking, tracking occlusion, and tracking loss.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] 1. The present invention provides a real-time, high-speed target recognition and tracking solution based on a domestically developed artificial intelligence framework, including artificial intelligence-based target recognition and tracking functions, as well as automatic target recognition-to-tracking functions. The solution fully designs and optimizes the deployment method during end-side deployment, combining the optimization of the Ascend CANN parallel asynchronous bus computing architecture for parallel computing. During implementation, the solution uses three different levels of underlying API interfaces based on the Ascend MindSpore framework to implement time-sharing switching deployment and scheduling for the target recognition and tracking processes. High-level APIs are used to strictly manage training and inference, perform mixed-schedule training, and perform tuning and debugging on the algorithm network, control the execution process of the entire network, and implement neural network training and inference. Second, mid-level APIs are used to flexibly construct neural networks and control the execution process, rapidly implementing model algorithm logic. Loss functions and optimization methods are added to the neural network model through the network ontology, LOSS function, and Optimizer interface. Finally, mathematical and logical operations such as tensor definition, operator calculation, and self-differentiation are implemented through the underlying APIs, greatly improving the intelligence of target recognition and tracking.

[0015] 2. This solution uses multi-threaded asynchronous neural network rapid scheduling to achieve concurrent and independent operation modes of the recognition network and tracking network. It has low inference time and is suitable for rapid deployment. It can simultaneously process multiple video streams of different spectral images and perform recognition and tracking network inference on each video stream. It can automatically identify targets and automatically switch to the continuous tracking process of the current target according to user needs without the need for additional operator intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the overall architecture diagram of the solution of the present invention;

[0017] Figure 2 This is a schematic diagram of multi-threaded pipeline scheduling in the present invention;

[0018] Figure 3 Schematic diagram of the software execution flow of the present invention;

[0019] Figure 4 This is a logic diagram of the automatic identification and tracking function flow of the present invention;

[0020] Figure 5 It is a rendering of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] See also Figure 1-5 The present invention provides a real-time, high-speed target recognition and tracking solution based on a domestically produced artificial intelligence framework. The technical solution is as follows: The AI ​​calculation of the present invention includes two network frameworks: target recognition and target tracking. Its core working logic is: 1. Use the recognition network to trigger the tracking network. When the recognition network recognizes a trusted target, it quickly switches to the tracking network to lock the target and subsequently accurately track it, predict the trajectory, etc. When the target is lost from the field of view, the tracking stops to prevent mistracking. At this time, the recognition network is reactivated to search for suspected targets in the field of view; 2. Manually select the target in the field of view, activate the tracking network to lock the target and subsequently accurately track it, predict the trajectory, etc. When the target is lost from the field of view, the tracking stops to prevent mistracking; 3. Self-developed long-term tracking algorithm with powerful anti-loss and re-capture functions, without any restrictions on target categories, truly achieving tracking of all things. The long-term tracker will perform in-depth feature extraction on the target during the tracking process, and has a persistent and controllable re-capture capability when the target disappears and re-enters the field of view, making up for the defect that the recognition algorithm can only re-capture targets of specific categories.

[0023] After the front-end multi-channel video images are input, they are switched in time by the target recognition and target tracking networks, which activate and schedule each other, saving computing resources and the power consumption of the network always running in the foreground. Calling the powerful hardware visual pre-processing core can quickly complete image processing operations such as cutout, scaling, overlay, splicing, histogram statistics, color remapping, format conversion, Remap transformation, filtering, etc. (the maximum resolution supported is 8192*8192, and the maximum processing time of a single operation at the maximum resolution is 5ms), which greatly reduces the time consumption of the pre-processing stage (the actual measurement of 1080P simultaneous cutout, scaling, filling and format conversion takes only 3ms). The pre-processed data is sent to the memory buffer of the NPU for subsequent inference. In the network inference part, this solution has made special optimizations to the training framework and underlying scheduling at the training layer and scheduling layer. The inference time is reduced by 20% compared with the non-optimized one, and the target recognition model The time consumption of the model is as follows (small model: 3-4ms, medium model: 7-8ms, large model: 12-13ms). The small model of the recognition algorithm of this solution has 3 times the number of parameters of the mainstream model used in HiSilicon 3599 / 3403, and the inference time is only 50% of that, which has a huge advantage in inference speed. After the inference is completed, it is output in two forms. Users can choose to directly transmit the inference results to the backend according to the needs of the backend platform, that is, the coordinates of the center point of the identified target and the coordinate sequence of the recognition box range (no time consumption). The system can also superimpose the inference results into the video stream encoding output (1ms). Therefore, the maximum total time consumption of the recognition network is 18ms, thanks to the pipeline scheduling of software threads, such as Figure 2 As shown in the figure, theoretically, when using a small model, it can meet the real-time inference process at a single-channel 1080P 50Hz (using a large model) -200Hz (using a small model) frame rate.

[0024] The tracking network is more complex than the recognition network. However, since the working logic of the present invention is that the recognition network activates the tracking network, the two networks are in a time-sharing serial working mode. The resources and time of each working module network are independent. The current pre-processing time of the tracking network is basically the same as that of the recognition network. After the pre-processing is completed, it is sent to the NPU for model inference. The present invention has a variety of inference algorithms. The algorithm time fluctuates according to the template size. The specific time consumption is as follows: 1. High-precision short-time tracking algorithm (small model: 1-3ms, Large model: 8ms-10ms. When the target is lost, it can only rely on the recognition module to recapture a specific category). 2. Long-term tracking algorithm (Small model: 1-3ms, Medium model: 4-6ms, Large model: 8-10ms. It has the ability to recapture after loss, is independent of the recognition algorithm, and can recapture any category). Both algorithms automatically update the target template during the tracking process (activating the update takes 1-2ms) and feature adaptive target template selection. After tracking inference, the output is in two forms. Depending on the needs of the backend platform, users can choose to directly transmit the inference results to the backend, namely the coordinates of the center point of the tracked target and the coordinate sequence of the tracking box range (which is time-saving). Alternatively, this solution can overlay the inference results into the video stream encoding output (1ms). Therefore, the total processing time of the tracking network is a maximum of 18ms, which can meet real-time inference requirements at a frame rate of 50Hz (large model) to 200Hz (small model) at 1080p resolution.

[0025] The software execution process of this solution is as follows Figure 3 As shown in the figure, the UART / ETH control thread continuously monitors the serial port and network port data, and is responsible for processing the sending and receiving tasks of the serial port or network port. The working logic of this module is as follows: 1. When the UART / ETH control thread receives an instruction, it will automatically parse the instruction, convert the instruction information into the corresponding tracker working status and synchronize it to the Decoder module. Subsequent video processing related modules perform corresponding processing based on the status information; 2. When the detector and tracker report detection and tracking information to the UART / ETH control thread, the UART / ETH control thread will process the reported information according to the instruction protocol and send it to the external master control immediately after processing.

[0026] The main thread of the recognition and tracking algorithm has different algorithm flows, as shown in the figure above, depending on the enabled functions. First, the Decoder module decodes the input video frame by frame. The decoded image is then recognized, tracked, or directly output according to the main control instructions. The processed image is overlaid with display information according to the protocol requirements. The OSD module outputs the image, which is then sent to the encoder for image encoding and sent to the external main control via the RTSP network video stream.

[0027] The target recognition thread algorithm mainly includes two threads: recognition pre-processing, target recognition, and recognition post-processing. The target recognition thread pre-processes the input current frame image to meet the requirements of the recognition network input, sends it to the recognition thread for model inference, and obtains the image output. The recognition post-processing thread is responsible for matching the target ID, determining the target category, and outputting the target recognition accuracy based on the network output and combined with information such as the time axis. If the target is lost and re-enters the field of view within 3 seconds, it can automatically match the previous ID. It also has the function of counting the number of targets of different categories.

[0028] The tracking algorithm thread is mainly responsible for processing the template extraction and calculation of the search area after the tracking instruction is issued (which can be issued by the external master control or by the recognition module). After determining the tracking template, the template and search area are sent to the tracking network for model reasoning, and the position and confidence of the tracked target are output. The main function of the tracking post-processing thread is to judge the current tracker status and dynamically update the tracking template. Its main logic is: if the tracking output confidence is greater than the set threshold, the tracking is considered normal; if the tracking output confidence is less than the set threshold, the target is considered lost or occluded. The occlusion loss processing logic is as follows:

[0029] 1. Occlusion

[0030] Trigger condition: The tracking confidence is less than the set threshold, the previous frame tracking is normal, and the target coordinates are not at the edge of the image. In this case, the target is considered to be in the occlusion state.

[0031] Search strategy: When the target is occluded, the system will continue to track and search at the target occlusion point until the target reappears, and then enter normal tracking mode. If the occlusion time is manually set, if the target does not appear after the occlusion time, it will automatically enter the target lost state and trigger the target lost search strategy.

[0032] 2. Lost

[0033] Trigger condition: The tracking confidence is less than the set threshold, and the target tracking output coordinates of the previous frame are at the edge of the image, in which case the target is set to be lost; the target occlusion timeout triggers the target lost state;

[0034] Search strategy: If the target is lost, the whole field of view will be searched for the target to determine whether the target reappears in the field of view. If the target reappears, the tracking state will be normal and the target tracking will continue. If the manually set target re-search time is exceeded, it is considered that the target will not reappear and the tracking will be automatically stopped.

[0035] After obtaining the tracking results, the tracking thread will immediately synchronize the miss distance information with the UART / ETH control thread. After synchronization, it will trigger the miss distance information output of the serial port and forward the miss distance information to the external master control. In addition, the tracking thread will send the tracking results to the post-processing thread, decide whether to update the template based on the user instructions and tracking results, and further pass the results to the OSD module for display information superposition. The OSD module is responsible for superimposing image frame sequence number, target identification related information, tracking miss distance information, turntable angle and other display information; the video encoding and streaming thread is responsible for sending the h264 or h265 encoding of the display image and pushing it to the video streaming server.

[0036] Automatic tracking detection Real-time automatic tracking function detects typical targets through recognition algorithms. After the recognition algorithm successfully captures the typical target, the tracking function is automatically triggered to achieve a complete automatic tracking closed loop. The function logic is as follows Figure 4 shown.

[0037] The real-time automatic tracking function triggers automatic tracking through the target detection network. First, the program will perform a status self-check after startup. After the self-check is completed, the recognition network will automatically start to identify the target in the picture. When a typical target appears in the picture, the recognition network will send the identified typical target coordinates to the tracking module. When the tracking module receives the target coordinates, it will start the tracking function.

[0038] The tracking algorithm of the present invention executes the following logic: after the tracker starts tracking, it will continue to judge the current tracking state. The specific tracking state is divided into the following three states:

[0039] ①Tracking is normal

[0040] When tracking is normal, the system will continue to track the target and output the tracking result miss amount in real time.

[0041] ②Tracking occlusion

[0042] When the tracked target is blocked by other targets, it enters the tracking blocked state and the tracker will continue to track the target in memory. The user can set the blocking timeout. When the blocking timeout is exceeded, the tracker will enter the blocking lost state and switch to the blocking lost tracking logic. If the tracker recaptures the target within the set blocking timeout, it will enter the tracking normal state. The system will continue to track the target and output the tracking result miss value in real time.

[0043] ③Tracking lost

[0044] When the tracked target enters the tracking loss state, the long-term tracking algorithm used by this system will perform a global recapture of the target. If the recapture timeout is not set, the system will continue to recapture the target until it reappears in the field of view. When the user sets the recapture timeout, the system will continue to perform a global recapture of the tracked target.

[0045] When the target is successfully recaptured, the system enters the normal tracking state. The system continues to track the target and outputs the tracking result miss value in real time.

[0046] When target recapture fails and the set recapture timeout period has not been exceeded, global recapture will continue.

[0047] When the target recapture fails and the set recapture timeout is exceeded, the recognition network will be re-triggered to search for a new typical target. When the typical target is recognized by the recognition network, the tracking logic will be re-entered.

[0048] In the description of the present invention, it should be understood that the terms "coaxial", "bottom", "one end", "top", "middle", "the other end", "upper", "one side", "top", "inside", "front", "center", "both ends", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0049] In addition, the terms "first", "second", "third" and "fourth" are used for descriptive purposes only and cannot be understood as indicating or suggesting relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first", "second", "third" and "fourth" may explicitly or implicitly include at least one such feature.

[0050] In the present invention, unless otherwise clearly stipulated and limited, the terms "install", "set", "connect", "fix", "screw" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integrated connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0051] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A real-time, high-speed target recognition and tracking method based on a domestically produced artificial intelligence framework, characterized by: It includes two network frameworks: target recognition and target tracking. The working logic is to use the recognition network to trigger the tracking network. When the recognition network recognizes a trusted target, it quickly switches to the tracking network to lock the target and subsequently accurately track and predict the trajectory. When the target is lost from the field of view, the tracking stops. At this time, the recognition network is reactivated to search for suspected targets in the field of view, manually select the target in the field of view, activate the tracking network, and start the long-term tracking algorithm. After the front-end multi-channel video image input, the target recognition and target tracking networks are divided into two networks for time-sharing switching, mutual activation and scheduling, and the hardware vision preprocessing core VPC is called to complete the cutout, scaling, overlay, splicing, histogram statistics, color remapping, format conversion, Remap transformation, and filtering image processing operations. The preprocessed data is sent to the memory buffer of the NPU The tracker will continue to track the target and output the target range coordinates in the target area and perform subsequent reasoning. After the reasoning is completed, it will be output in two forms, namely the coordinates of the center point of the identified target and the coordinate sequence of the identification frame range. The reasoning results are superimposed into the video stream encoding output. The execution logic of the long-term tracking algorithm is as follows. After the tracker starts tracking, it will continue to judge the current tracking state. The tracking state is divided into the following three types: Normal tracking: Continuously track the target and output the tracking result miss amount in real time; Tracking occlusion: When the tracked target is occluded by other targets, it enters the tracking occlusion state, and the tracker will continue to memory track. If the user sets the occlusion timeout, when the set occlusion timeout is exceeded, the tracker will enter the occlusion loss state and switch to the occlusion loss tracking logic. If the tracker recaptures the target within the set occlusion timeout, it will enter the normal tracking state. Tracking loss: When the tracked target enters the tracking loss state, the long-term tracking algorithm will globally recapture the target. If the recapture timeout is not set, the algorithm will continue to recapture until the target reappears in the field of view. When the user sets the recapture timeout, the algorithm will continue to globally recapture the tracked target. When the target is successfully recaptured, it will enter the normal tracking state. When the target recapture fails and does not exceed the set recapture timeout, it will continue to globally recapture. When the target recapture fails and exceeds the set recapture timeout, the recognition network will be re-triggered to search for a new target. When the target is recognized by the recognition network, it will re-enter the tracking logic. The target recognition and target tracking networks are in time-sharing serial working mode, and the resources and time of each working module network are independent.

2. The real-time high-speed target recognition and tracking method based on a domestic artificial intelligence framework according to claim 1 is characterized in that: The software execution process of this target recognition and tracking method is that the UART / ETH control thread continuously monitors serial port and network port data and is responsible for processing the sending and receiving tasks of the serial port or network port. When the UART / ETH control thread receives an instruction, it will automatically parse the instruction, convert the instruction information into the corresponding tracker working status and synchronize it to the Decoder module. Subsequent video processing related modules perform corresponding processing based on the status information. When the detector and tracker report detection and tracking information to the UART / ETH control thread, the UART / ETH control thread will process the reported information according to the instruction protocol and send it to the external master control immediately after processing.

3. The real-time high-speed target recognition and tracking method based on a domestic artificial intelligence framework according to claim 1 is characterized in that: First, the Decoder module decodes the input video frame by frame. The decoded image is identified, tracked, or directly output according to the main control instructions. The processed image is superimposed with display information according to the protocol requirements. The output image of the OSD module is sent to the encoder for image encoding and sent to the external main control through the RTSP network video stream.

4. The real-time high-speed target recognition and tracking method based on a domestic artificial intelligence framework according to claim 3 is characterized in that: The target recognition includes three threads: recognition pre-processing, target recognition and recognition post-processing. The target recognition thread pre-processes the input current frame image to meet the requirements of the recognition network input, sends it to the recognition thread for model inference, and obtains the image output. The recognition post-processing thread is responsible for matching the target ID, determining the target category, and outputting the target recognition accuracy function based on the network output and the timeline information. If the target re-enters the field of view within 3 seconds after being lost, the previous ID can be automatically matched. At the same time, it has the function of counting the number of targets of different categories.

5. The real-time high-speed target recognition and tracking method based on a domestic artificial intelligence framework according to claim 1 is characterized in that: The target tracking is responsible for processing the template extraction and calculating the search area after the tracking instruction is issued. After determining the tracking template, the template and search area are sent to the tracking network for model reasoning, and the position of the tracked target and the confidence of the tracked target are output.

6. The real-time high-speed target recognition and tracking method based on a domestic artificial intelligence framework according to claim 1, characterized in that: The target tracking also includes the function of judging the current tracker state and dynamically updating the tracking template. Its main logic is: if the tracking output confidence is greater than the set threshold, the tracking is considered normal; if the tracking output confidence is less than the set threshold, the target is considered lost or blocked.

Images