Artificial intelligence driven machine vision dynamic target intelligent tracking detection system

Abstract

The application discloses an artificial intelligence driven machine vision dynamic target intelligent tracking detection system and belongs to the technical field of machine vision. The processor is configured to: extract motion features of a target historical trajectory by using a time sequence encoder trained by a meta-learning; maintain an online adjustable mode embedding vector for the target; output a predicted position and a predicted uncertainty measure by gate soft fusion of at least two predictors based on the motion features and the mode embedding vector; store learned mode embedding vectors and corresponding trajectory features in a mode memory bank; trigger online adaptation according to a prediction error; perform data association based on the uncertainty measure and adjust the allocation of computing resources according to the importance of the target. Through the autonomous mode discovery driven by meta-learning, the gate soft fusion, the mode memory bank and the online adaptation mechanism, rapid adaptation and zero sample reuse to any motion mode are realized, and the tracking precision and real-time performance are improved.

Description

Technical Field

[0001] This invention relates to an artificial intelligence-driven machine vision dynamic target intelligent tracking and detection system, belonging to the field of machine vision technology. Background Technology

[0002] Dynamic target tracking is a core technology in machine vision, widely used in intelligent monitoring, autonomous driving, and other fields. Existing methods, such as DeepSORT, employ Kalman filtering and Hungarian matching architectures, perform well in uniform motion scenarios. However, when the target's motion pattern changes significantly, these methods may exhibit significant deviations in trajectory estimation due to motion model mismatch, posing a potential risk to driving safety.

[0003] In recent years, researchers have proposed methods such as interactive multi-model, particle filtering, and LSTM trajectory prediction to enhance motion pattern adaptability. Recently published patent document CN119904485B proposes a multi-target tracking method based on weak cues and trajectory prediction. This scheme uses LSTM to predict the target's future trajectory when occlusion occurs, and Kalman filtering for trajectory prediction when there is no occlusion, thus improving tracking performance in occluded scenes to some extent. However, this scheme still has the following inherent drawbacks: This scheme uses occlusion indicators for mode switching—it immediately switches to LSTM prediction when occlusion occurs and switches back to Kalman filtering immediately when occlusion ends. This binary hard-switching mechanism causes abrupt changes in the prediction model during mode switching, resulting in discontinuous trajectory estimation. Taking an autonomous vehicle gradually approaching the main road from a ramp as an example, the vehicle's motion mode gradually transitions from constant speed turning to straight-line acceleration. The handover between the two predictors should be completed with continuous weights, but the hard switch causes a jump in the prediction output at the handover point, affecting the decision stability of the downstream control module.

[0004] The occlusion state needs to be predefined and labeled, limiting the system's adaptability to preset modes. It cannot autonomously detect other motion modes, such as sharp turns, serpentine movements, or random changes of direction. When the target's motion characteristics exceed the preset range, the system performance degrades significantly.

[0005] Once the model is trained, its parameters are fixed, making it unable to adapt to newly emerging motion patterns during runtime and unable to continuously optimize based on trajectory data accumulated during actual operation.

[0006] Each time a similar movement pattern is encountered, adaptation must be started from scratch, and historical experience cannot be used to speed up the adaptation process.

[0007] Maintaining the same computational load for all targets makes real-time tracking difficult to achieve on resource-constrained edge devices.

[0008] It is evident that existing predictive models for dynamic target tracking suffer from the following systemic technical contradictions: incompatibility between fixed parameter models and changing motion patterns, conflict between hard switching mechanisms and the smoothness of continuous trajectories, and imbalance between limited computational resources and the demand for high-precision real-time tracking. Summary of the Invention

[0009] The purpose of this invention is to provide an artificial intelligence-driven machine vision dynamic target intelligent tracking and detection system. Through mechanisms such as meta-learning-driven autonomous pattern discovery, internalized dual-head predictor, gated soft fusion, pattern memory library, and online adaptive mechanism, the system can autonomously identify and adapt to any motion mode, significantly improving tracking accuracy and robustness while maintaining real-time performance.

[0010] To achieve the above objectives, the present invention provides an artificial intelligence-driven machine vision dynamic target intelligent tracking and detection system, including a processor and a memory.

[0011] The processor is configured to: The system acquires continuous video frames and detects targets; extracts motion features of the target's historical trajectory using a temporal encoder trained by eurythmics; maintains online adjustable motion pattern representation parameters for the target; based on the motion features and the motion pattern representation parameters, outputs the predicted position through adaptive fusion of at least two predictors, where the fusion weights are dynamically allocated by a gating network according to the current motion state; outputs a measurement of the prediction uncertainty; stores the learned motion pattern representation parameters and corresponding trajectory features in a motion pattern memory, and supports retrieval matching to achieve zero-sample reuse; based on the prediction error trigger, quickly adapts to the current motion pattern by adjusting the motion pattern representation parameters and the parameters of the gating network; performs data association based on the prediction uncertainty measurement, and adjusts the allocation of computing resources according to the importance of the target; outputs the tracking result.

[0012] The memory is configured to store the motion pattern memory and computer program.

[0013] Furthermore, the temporal encoder is at least one of a long short-term memory network, a gated recurrent unit network, or a Transformer encoder; the meta-training employs a model-independent meta-learning algorithm, which is trained by constructing a meta-task consisting of a support set and a query set.

[0014] Furthermore, the at least two predictors include a linear motion prediction head and a nonlinear motion prediction head; the linear motion prediction head employs a linear fully connected layer, and the nonlinear motion prediction head employs a multilayer fully connected network with a nonlinear activation function.

[0015] Furthermore, the uncertainty measure of the prediction adopts a mixture Gaussian distribution parameter, including the weights, mean, and standard deviation of multiple Gaussian components; the data association based on the uncertainty measure includes calculating Mahalanobis distance or weighted Euclidean distance according to the uncertainty measure.

[0016] Furthermore, the method of adjusting the allocation of computational resources based on target importance includes: determining a target importance score based on at least one of the following factors: target size, movement speed, occlusion status, and distance from the region of interest; detecting high-importance targets at a higher frequency and low-importance targets at a lower frequency.

[0017] Furthermore, the processor is also configured to: when a new target appears, accumulate trajectory data for an initial preset number of frames using Kalman filtering, and initialize the motion mode representation parameters to default values.

[0018] Furthermore, the storage strategy of the motion pattern memory bank includes: after completing an online adaptation, storing the updated motion pattern representation parameters and their trajectory fingerprints into the motion pattern memory bank; when searching for a match, calculating the cosine similarity or Euclidean distance between the current trajectory fingerprint and each entry in the bank; if the maximum similarity exceeds a preset threshold, directly reusing the corresponding motion pattern representation parameters.

[0019] Furthermore, when the processor adjusts the motion mode representation parameters and the parameters of the gating network, it uses the historical trajectory of the most recent preset number of frames as the support set, performs gradient descent for a preset number of steps, and does not update the parameters of the temporal encoder.

[0020] Furthermore, the processor is also configured to: when adaptive adjustment is triggered based on prediction error, prioritize searching the motion pattern memory; if a similar pattern is found, it is directly reused to achieve zero-sample adaptation; if no similar pattern is found, the adjustment step is executed, and the adjusted motion pattern representation parameters are stored in the motion pattern memory.

[0021] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the functions of the above-described system.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: The system does not require manual preset of motion state categories and can automatically discover and adapt to any motion mode through meta-learning, significantly improving tracking robustness in complex scenarios. It only requires 5 frames of trajectory data and 5 steps of gradient update to adapt to new motion patterns, which improves the adaptation efficiency by about 4-6 times compared with traditional fine-tuning; when the pattern memory is hit, zero-sample adaptation is achieved. The pattern memory library enables the system to accumulate experience in motion patterns. As the running time increases, the system's speed of adapting to common motion patterns gradually improves, achieving continuous performance evolution. Gated soft fusion replaces the hard switching mechanism of traditional methods, making the predicted trajectory smoother. In the case of gradual motion, the two predictors are output with continuous weights to avoid abrupt changes in the predicted output. Online adaptation updates only motion pattern representation parameters and gating network parameters, resulting in extremely low computational overhead; adaptive detection frequency scheduling enables the system to run in real time on edge devices. By quantifying the uncertainty of prediction through Gaussian mixture output and directly using it for data association, the robustness of occlusion recovery and multi-target tracking is improved. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the architecture of the AI-driven machine vision dynamic target intelligent tracking and detection system of the present invention; Figure 2 This is a flowchart illustrating the meta-training phase of the present invention. Figure 3 This is a flowchart illustrating the online reasoning and adaptive phases of the present invention; Figure 4 This is a schematic diagram of the internalized dual-head predictor and gated soft fusion structure of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] For ease of understanding, the following provides a unified explanation of several key terms appearing in this embodiment: A temporal encoder: a recurrent neural network used to extract motion features from the historical trajectory of a target. In this embodiment, a Long Short-Term Memory (LSTM) network is used, but it can also be replaced by a gated recurrent unit network (GRU) or a Transformer encoder.

[0027] Pattern embedding vector: A low-dimensional parameter vector that can be adjusted online to represent the motion pattern features of a target. The ability to rapidly adapt through meta-learning is one of the core parameters for achieving online adaptation in this invention.

[0028] Gated networks: a lightweight neural network that dynamically calculates the fusion weights of each predictor based on the current motion features and pattern embedding vectors, enabling soft switching of expert predictions.

[0029] Motion pattern memory: A data structure stored in memory to save learned pattern embedding vectors and corresponding trajectory fingerprints. It supports trajectory feature-based retrieval and matching, enabling zero-sample reuse.

[0030] Meta-learning: A training paradigm that enables models to learn how to learn. This invention employs the Model-Independent Meta-Learning (MAML) algorithm, which, through training on a large number of meta-tasks, enables the temporal encoder to acquire the meta-ability to quickly adapt from a small number of new trajectories.

[0031] Please see Figure 1 The AI-driven machine vision dynamic target intelligent tracking and detection system provided in this embodiment of the invention includes a processor and a memory.

[0032] The processor is configured to execute core algorithms such as target detection, trajectory prediction, pattern discovery, online adaptation, data association, and resource scheduling. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), such as an embedded AI chip from the NVIDIA Jetson series, or a combination thereof.

[0033] The memory is communicatively connected to the processor and configured to store motion pattern memory and computer programs. The memory may be a read-only memory (ROM), random access memory (RAM), flash memory, solid-state drive (SSD), or a combination thereof.

[0034] Optionally, the system also includes an image acquisition module configured to acquire continuous video frames. The image acquisition module is communicatively connected to the processor and transmits the acquired video frames to the processor for processing.

[0035] Optionally, the system also includes an output module configured to output tracking results, including tracking boxes, target IDs, and motion state labels, for use by downstream systems. In this embodiment, the downstream system can be an autonomous driving decision module or a monitoring platform.

[0036] In this embodiment, the model acquires the meta-learning ability through meta-training during the offline phase. This meta-training process is executed on an offline computing platform, and the model parameters are deployed to the target computing platform after training is completed.

[0037] For details on meta-task construction, please refer to [link / reference]. Figure 2 During the offline training phase, the processor extracts a large number of trajectory segments from the trajectory dataset, each segment being 15 frames long. Each segment is divided into a support set and a query set. Each segment is considered a meta-task, representing a potential motion pattern. In this embodiment, approximately 200,000 meta-tasks can be constructed, covering various patterns such as linear motion, constant speed turning, sharp turns, serpentine motion, spiral motion, and sudden stop-start.

[0038] Regarding the structure and selection of the temporal encoder, the processor uses a single-layer LSTM as the temporal encoder (64 hidden units). The input is the historical trajectory sequence, i.e., the position coordinates of the most recent 10 frames. Each frame contains x and y coordinates, for a total of 20 values. The output is the hidden state at the current moment (a 64-dimensional vector). This LSTM encoder is a shared meta-parameter for all meta-tasks.

[0039] Specifically, the choice of a single-layer 64-cell LSTM is based on the following considerations: LSTM has a forget gate and an input gate, which can effectively alleviate the gradient vanishing problem and perform more stably when processing medium-length time series data such as trajectory sequences. Compared to GRU, LSTM has independent forget gate and input gate, which is advantageous in scenarios where trajectory feature information needs to be retained for a long time to assist in pattern embedding vector learning. The meta-training phase requires training multiple meta-tasks simultaneously. Excessive parameter count can increase computational overhead. A single-layer 64-unit LSTM achieves the optimal balance between prediction accuracy and training efficiency. In actual testing on an embedded computing platform, the LSTM encoder's single forward inference time is less than 0.2 milliseconds, meeting the real-time tracking requirement of more than 30 frames per second.

[0040] Understandably, using multi-layer LSTM or networks with larger hidden dimensions can improve performance, but at the cost of inference speed. Using GRU networks can reduce the number of parameters, but may result in a loss of the ability to model long-term trajectory dependencies. All of the above alternatives fall within the scope of protection of this invention.

[0041] Furthermore, the processor associates a 16-dimensional learnable pattern embedding vector with each meta-task to capture the motion pattern features of that task. At the start of meta-training, the pattern embedding vector is randomly initialized; during meta-training, it is optimized to quickly adapt to the meta-parameters of the specific task through a small number of gradient updates.

[0042] Further, please refer to Figure 4 The processor concatenates the hidden state (64-dimensional) with the 16-dimensional pattern embedding vector to form an 80-dimensional vector, which is then input into the linear motion prediction head and the nonlinear motion prediction head, respectively.

[0043] Specifically, the linear motion prediction head uses a fully connected linear layer to output the predicted position offset. It excels at predicting linear motion patterns such as uniform speed or uniform acceleration.

[0044] Specifically, the nonlinear motion prediction head uses a multi-layer fully connected network with ReLU activation, which is good at predicting nonlinear motion modes such as turning, random change of direction, and occlusion.

[0045] Simultaneously, the gating network outputs fusion weights for the two prediction heads based on the same 80-dimensional concatenated vector, and these fusion weights are normalized using the Softmax function. The final predicted position is the weighted sum of the two expert outputs.

[0046] Understandably, the number of the at least two predictors is not limited to two, and can be expanded to three or more predictors depending on the complexity of the motion pattern; the internal structure of the predictors can also be replaced with other forms such as adaptive filters. The weighted fusion is a continuous weight allocation, which is different from the binary hard switching mechanism in the prior art.

[0047] Furthermore, the processor branches out from the 80-dimensional concatenated vector, outputting a Gaussian mixture distribution parameter with three Gaussian components. Each Gaussian component includes: weights, x-direction mean, y-direction mean, x-direction standard deviation, and y-direction standard deviation. Specifically, the weights are normalized using Softmax, the x-direction mean is limited to the range [-1, 1] using the Tanh function, the y-direction mean is limited to the range [-1, 1] using the Tanh function, and the x-direction standard deviation is guaranteed to be positive using the Softplus function. The Gaussian mixture output represents the quantification of the uncertainty of the predicted position.

[0048] Furthermore, regarding the meta-training loss function, for each meta-task, the processor first performs a 5-step inner loop gradient update on the support set, updating only the pattern embedding vectors and the parameters of the gated network, while fixing the temporal encoder and prediction head. After the inner loop, the total loss is calculated on the query set, including: State classification loss: The cross-entropy loss function is used to determine the state of the target, such as whether it exists or is occluded, by using a classification branch.

[0049] Negative log-likelihood loss of Gaussian mixture: For the true trajectory points in the query set, calculate their negative log-likelihood under the Gaussian mixture distribution, which helps the model learn the accurate prediction distribution.

[0050] The total loss on the query set is used to update the meta-parameters, namely the initial weights of the sequential encoder, prediction head, and gating network. Through iterative training on numerous meta-tasks, the meta-parameters are brought into a state that allows them to quickly adapt to new motion patterns.

[0051] In this embodiment, the meta-training uses the following hyperparameter settings: the optimizer is Adam, the inner loop learning rate is 0.01, the outer loop learning rate is 0.001, the inner loop step count is 5, the meta-batch size is 32, and the number of meta-training epochs is 100. The learning rate uses a cosine annealing decay strategy, initially 0.001, decaying to 0.9 times its original value every 10 epochs. To prevent overfitting, a dropout rate of 0.2 is used for regularization of the temporal encoder and prediction head. The above hyperparameters are preferred values ​​for this embodiment; those skilled in the art can adjust these parameters according to specific application scenarios, all of which fall within the protection scope of this invention.

[0052] In this embodiment, after the meta-training is completed, the processor retains a blank motion pattern memory in the memory, the capacity of which is configurable. In this embodiment, it can be configured to store 50 patterns.

[0053] Specifically, regarding trajectory fingerprint extraction, for the motion pattern representation parameters (i.e., pattern embedding vectors) obtained after each online adaptation, the processor simultaneously calculates the feature fingerprint of the corresponding trajectory segment. In a preferred embodiment, the average value of the hidden states output by the temporal encoder or the hidden state at the last moment is used as the fingerprint. This fingerprint is used for subsequent pattern retrieval and matching.

[0054] Regarding the storage strategy, after completing an online adaptation, the processor stores the motion pattern characterization parameters and their fingerprints into the motion pattern memory. If the fingerprint is highly similar to an existing entry, an exponential moving average is used to merge and update the entry, and the count of that pattern is increased; otherwise, it is added as a new entry. When the memory is full, the least frequently used pattern is discarded.

[0055] Understandably, the similarity threshold can be adjusted according to the specific application scenario; alternatively, other similarity measures such as Euclidean distance can be used to replace cosine similarity.

[0056] Regarding retrieval reuse, in subsequent tracking, for the current trajectory segment, the processor first extracts its fingerprint and calculates its similarity with the fingerprints of each entry in the motion pattern memory. If the maximum similarity exceeds a preset threshold, the corresponding motion pattern representation parameters are directly retrieved as the initial values ​​for the current pattern embedding, without needing to perform inner loop gradient updates, thus achieving zero-shot adaptation. If no match is found, the inner loop adaptation is triggered according to the normal process.

[0057] Please see Figure 3 In this embodiment, the processor in the online inference phase performs the following operations: When a new target appears, the processor uses Kalman filtering for cold start to accumulate trajectory data from the previous 10 frames and initializes the 16-dimensional pattern embedding vector as a zero vector.

[0058] For targets that have accumulated trajectories of 10 or more frames, the processor executes the following sub-steps: Specifically, regarding feature extraction, the processor extracts the position coordinates of the most recent 10 frames from the target trajectory, inputs them into the temporal encoder, and obtains the hidden state.

[0059] Specifically, regarding memory retrieval, the processor searches the motion pattern memory. If a similar pattern is found, the motion pattern representation parameters in the memory are used directly; otherwise, the currently saved pattern embedding or the initial default value is used.

[0060] Specifically, regarding prediction and uncertainty output, the processor embeds and concatenates the latent state and pattern, calculates expert weights through a gating network, outputs the predicted position offset through linear motion prediction heads and nonlinear motion prediction heads, and outputs uncertainty parameters through a Gaussian mixture layer.

[0061] Specifically, regarding data association, the processor performs Mahalanobis distance matching between the predicted location and the detection results of the current frame, and combines appearance features to complete the data association.

[0062] Specifically, regarding resource scheduling, the processor dynamically adjusts the frame interval for the next detection based on the importance of the target. High-importance targets are detected every frame, while low-importance targets can be detected every other frame, in order to optimize the computational load.

[0063] Furthermore, regarding error monitoring and adaptive triggering, the processor calculates the average Euclidean distance between the predicted and actual detected positions of the target over a past period after processing a preset number of frames. If the average distance exceeds a preset threshold, it is considered that the current motion mode characterization parameters are no longer suitable for the target's motion characteristics, triggering the adaptive process. Zero-shot priority: Extract fingerprints using the current trajectory segment and search the motion pattern memory again. If a match is found, directly replace the motion pattern representation parameters without gradient updates.

[0064] Fast inner loop adaptation: If a match is missed, the most recent 5 frames of trajectory are used as the support set, and 5 steps of gradient descent are performed. Only the motion mode representation parameters and gating network parameters are updated, without updating the temporal encoder. The updated motion mode representation parameters are stored in the motion mode memory.

[0065] Post-adaptation smoothing: The adapted pattern embedding is subjected to an exponential moving average with the old value to avoid prediction jumps caused by abrupt pattern changes.

[0066] It is understood that the values ​​in the above steps are preferred values ​​in this embodiment. Those skilled in the art can adjust these parameters according to specific application scenarios, and all of these adjustments will fall within the protection scope of this invention.

[0067] Regarding data association based on uncertainty perception, the processor utilizes the Gaussian mixture parameters output by the prediction module and employs Mahalanobis distance as the metric during data association. Mahalanobis distance considers the covariance matrix and measures the statistical difference between the predicted location and the detection result. Compared to Euclidean distance, Mahalanobis distance tolerates greater location deviation when prediction uncertainty is high, and requires the detection result to be closer to the predicted location when uncertainty is low, thereby improving the robustness of the association.

[0068] It is understood that the distance metric is not limited to Mahalanobis distance, but may also use other distance metrics that take into account uncertainty, such as weighted Euclidean distance.

[0069] Regarding adaptive detection frequency scheduling, the processor maintains an importance score for each target. This importance score is calculated based on factors such as target size, movement speed, occlusion status, and distance to the region of interest. High-importance targets are detected at a higher frequency, while low-importance targets are detected at a lower frequency. This reduces computational load while maintaining tracking quality, enabling the system to achieve real-time performance on edge devices. Example 1

[0070] In this embodiment, the system is deployed on the computing platform of a Level 2+ autonomous vehicle. An exemplary hardware configuration is an NVIDIA Jetson AGX Xavier embedded computing platform, which uses a 1080p resolution, 30 frames per second forward-facing camera to capture dynamic targets such as other vehicles and pedestrians in front of the vehicle in real time.

[0071] Regarding the meta-training phase, approximately 200,000 trajectory segments were extracted from the MOT17, KITTI Tracking, and synthetic trajectory datasets, with each segment consisting of 15 frames. These segments were divided into training, validation, and test sets in an 8:1:1 ratio. Meta-training was performed using the aforementioned hyperparameters. After approximately 100 epochs, the initial parameters of the LSTM encoder were optimized to a state that facilitates rapid adaptation to new motion patterns. After meta-training, the model's meta-parameters were saved for subsequent online inference.

[0072] Online reasoning stage: Video frames are input to the processor at 30 frames per second. The processor runs the YOLOv8n detection module, which detects close-range targets every frame and long-range stable targets every other frame.

[0073] For each tracked target, the processor maintains a queue of its historical trajectories.

[0074] When the target is first created, the processor uses a Kalman filter to cold start and accumulate the first 10 frames of data, and initializes the 16-dimensional pattern embedding as a zero vector.

[0075] After entering steady-state tracking, for targets that have accumulated more than 10 frames of trajectory, the aforementioned online inference and adaptive process is executed.

[0076] Deployment Results: In real-world testing under comprehensive urban road conditions in autonomous driving scenarios, facing complex dynamic scenarios such as sudden lane changes and pedestrians crossing the road, the system operated continuously for 24 hours, achieving a tracking accuracy of 92.3%, with an average of less than 5 identity switching times per thousand frames and a stable frame rate of over 30 frames per second. The online adaptive mechanism was successfully triggered multiple times when encountering sudden target changes and occlusion scenarios, with prediction accuracy returning to normal levels after an average adaptation time of approximately 5 frames of trajectory data. After accumulating 50 common motion patterns in the pattern memory, the adaptation time for new motion patterns decreased from an average of 5 steps of gradient updates to zero-sample reuse, validating the experience accumulation and continuous evolution capabilities of this invention. Example 2

[0077] In this embodiment, the processor tracks a vehicle traveling at a constant speed on a highway, establishes a corresponding 16-dimensional pattern embedding vector for it, and stores it in a pattern memory. When another vehicle of the same model and in the same scene enters the field of view, the processor extracts its trajectory fingerprint. Similarity calculations show a similarity greater than 0.8 with the uniform linear motion pattern, so the processor directly reuses the 16-dimensional pattern embedding vector from the memory as the initial value. At this point, without performing any gradient updates, the predictor can output accurate prediction results, achieving zero-shot rapid adaptation.

[0078] This invention is not limited to the above embodiments. The following alternatives also fall within the protection scope of this invention. Those skilled in the art can choose the appropriate alternative according to the specific application scenario.

[0079] Regarding the replacement of the temporal encoder, the embodiment uses a single-layer LSTM with 64 hidden units, which can be replaced by a multi-layer LSTM, a bidirectional LSTM, a GRU network, or a Transformer encoder. Using a GRU network reduces the number of parameters, making it suitable for edge devices with more demanding computing requirements; using a Transformer encoder improves prediction accuracy but reduces real-time performance. Regardless of the temporal encoder used, the core function of extracting motion features from historical trajectories is achieved.

[0080] Regarding the replacement of the pattern embedding vector dimension, the embodiment uses a 16-dimensional pattern embedding vector, which can be adjusted to 8-dimensional for scenarios with simpler motion patterns or 32-dimensional for scenarios with more complex motion patterns, depending on the task complexity. Both can achieve the function of capturing motion pattern features.

[0081] Regarding alternatives to the number of predictors, the embodiment uses two predictors. In alternatives, this can be expanded to three or more predictors depending on the complexity of the motion pattern, for example, by adding a sharp-turn motion predictor, a random motion predictor, etc.

[0082] Regarding alternative prediction head structures, the linear motion prediction head in this embodiment employs a fully connected layer, while the nonlinear motion prediction head employs a multi-layer fully connected network. In alternative solutions, the nonlinear prediction head can adopt an adaptive filter-based structure to enhance its adaptability to different nonlinear modes.

[0083] Regarding alternative expressions for prediction uncertainty, the example uses a Gaussian mixture distribution with three Gaussian components to represent the prediction uncertainty. Alternative solutions may employ a Laplace distribution, a discretized distribution, or other probability distributions, all of which quantify the prediction uncertainty.

[0084] Regarding the substitution of memory capacity and matching threshold, in this example, the memory capacity is 50 and the cosine similarity threshold is 0.8. In alternative solutions, the capacity can be increased to 100 or more to accommodate richer scene patterns; the threshold can be adjusted according to the application scenario. A higher threshold results in higher matching accuracy but lower recall, while a lower threshold results in higher recall but may lead to false matches.

[0085] Regarding the alternative to updating the number of steps and learning rate in the inner loop, in the example, the number of steps in the inner loop is 5 and the learning rate is 1e-5. In the alternative, the number of steps can be adjusted in the range of 3 to 10. Specifically, the more steps, the higher the adaptation accuracy, but the greater the computational cost. The learning rate can be adjusted in the range of 1e-6 to 1e-4.

[0086] Regarding the alternative to the adaptive trigger threshold, in this embodiment, the error monitoring threshold is 10 pixels, which can be adjusted according to the image resolution and target size of the specific scene.

[0087] Regarding the alternatives to the meta-training optimizer and the learning rate decay strategy, the embodiment uses the Adam optimizer and the cosine annealing decay strategy. Alternatives include the use of a stochastic gradient descent (SGD) optimizer, which may converge more stably when the outer loop learning rate is small; a step decay strategy; or other adaptive optimizers, all of which can achieve convergent optimization of the meta-parameters.

[0088] Regarding alternative detection frequency scheduling strategies, adaptive detection frequency scheduling based on target importance is one of the original features of this invention. In the alternative, the importance score can incorporate more factors such as the number of historical losses and tracking confidence; the detection frame interval can be set according to a step function or a continuous function.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-driven machine vision dynamic target intelligent tracking and detection system, characterized in that, include: The processor is configured as follows: Acquire consecutive video frames and detect targets; Motion features of the target's historical trajectory are extracted using a temporal encoder obtained through econometric training. To maintain online adjustable motion pattern characterization parameters for the target; Based on the motion features and the motion pattern characterization parameters, the predicted position is output through adaptive fusion of at least two predictors, wherein the fusion weights are dynamically allocated by the gating network according to the current motion state. Output a measure of the uncertainty of the forecast; The learned motion pattern representation parameters and corresponding trajectory features are stored in a motion pattern memory, and retrieval matching is supported to achieve zero-sample reuse. Triggered by prediction error, the motion mode characterization parameters and the gating network parameters are adjusted to quickly adapt to the current motion mode. Data association is performed based on the uncertainty measure of the prediction, and the allocation of computing resources is adjusted according to the importance of the target; Output the tracking results; The memory is configured to store the motion pattern memory and computer program.

2. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The temporal encoder is at least one of a long short-term memory network, a gated recurrent unit network, or a Transformer encoder; the meta-training adopts a model-independent meta-learning algorithm, which is trained by constructing a meta-task consisting of a support set and a query set.

3. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The at least two predictors include a linear motion prediction head and a nonlinear motion prediction head; the linear motion prediction head employs a linear fully connected layer, and the nonlinear motion prediction head employs a multilayer fully connected network with a nonlinear activation function.

4. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The uncertainty measure of the prediction adopts a mixture Gaussian distribution parameter, including the weights, mean and standard deviation of multiple Gaussian components; the data association based on the uncertainty measure includes calculating Mahalanobis distance or weighted Euclidean distance according to the uncertainty measure.

5. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The method of adjusting computational resource allocation based on target importance includes: determining a target importance score based on at least one of the following factors: target size, movement speed, occlusion status, and distance from the region of interest; detecting high-importance targets at a higher frequency and low-importance targets at a lower frequency.

6. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The processor is also configured to: when a new target appears, accumulate trajectory data for an initial preset number of frames using Kalman filtering, and initialize the motion mode characterization parameters to default values.

7. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The storage strategy of the motion pattern memory bank includes: after completing an online adaptation, the updated motion pattern representation parameters and their trajectory fingerprints are stored in the motion pattern memory bank; when searching for a match, the cosine similarity or Euclidean distance between the current trajectory fingerprint and each entry in the bank is calculated; if the maximum similarity exceeds a preset threshold, the corresponding motion pattern representation parameters are directly reused.

8. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, When the processor adjusts the motion mode representation parameters and the parameters of the gating network, it uses the historical trajectory of the most recent preset frame number as the support set, performs gradient descent for a preset number of steps, and does not update the parameters of the temporal encoder.

9. The AI-driven machine vision dynamic target intelligent tracking and detection system according to claim 1, characterized in that, The processor is further configured to: when adaptive adjustment is triggered based on prediction error, prioritize searching the motion pattern memory; if a similar pattern is found, it is directly reused to achieve zero-sample adaptation; if no similar pattern is found, the adjustment step is executed, and the adjusted motion pattern representation parameters are stored in the motion pattern memory.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the functions of the system according to any one of claims 1 to 9.

Images