Target tracking method, device and equipment based on AI visual identification

Abstract

The invention relates to the technical field of image data processing, and provides a target tracking method, device and equipment based on AI visual identification, and the method comprises the steps: tracking a target in real time through satellite positioning, obtaining the image feature information of the target and the map position information of the target during missing tracking when the target is not tracked, and generating the associated feature information; inputting the associated feature information into a preset motion prediction model, outputting motion prediction information of the target, and performing multi-size feature analysis on the associated feature information to obtain multi-scale feature information; and fusing the motion prediction information and the multi-scale feature information, and inputting a preset AI visual identification model to obtain real-time position information and motion track information of the target. The data are analyzed through an AI visual identification model, the real-time position information and the motion trail information of the target are output, and the problem that the target is lost and tracking fails when the target is in complex environments such as sheltered, illumination changing or rapid motion is solved.

Description

Technical Field

[0001] The present application relates to the technical field of image data processing, and in particular to target tracking methods, devices and equipment based on AI visual recognition. Background Art

[0002] In modern society, with the rapid development of science and technology, target tracking technology has been widely used in many fields, including intelligent transportation, unmanned driving, security monitoring, etc. By using artificial intelligence and machine learning technology, the accuracy and reliability of the target tracking system have been greatly improved. It can monitor and identify the movement trajectory of the target in real time, and warn of abnormal behavior, providing more intelligent solutions for all walks of life.

[0003] Among the related technical means, target tracking technology mainly relies on video surveillance systems, which capture the image of the target through fixed or mobile cameras, and identify and track the target through image processing algorithms. Traditional methods usually include background subtraction, optical flow method and Kalman filtering, which determine the location information of the target by analyzing the image changes in consecutive frames, thereby achieving continuous tracking of the target.

[0004] Regarding the above technical solution, although basic tracking of the target can be achieved through the existing target tracking technology, in actual applications, when the target is in a complex environment such as being blocked, with changing lighting or moving rapidly, there is a problem of target loss, resulting in tracking failure. Summary of the invention

[0005] In order to improve the problem of target loss and tracking failure when the target is in complex environments such as occlusion, lighting changes or rapid movement, this application provides a target tracking method, device and equipment based on AI visual recognition. Through satellite positioning and multi-feature information fusion, accurate tracking and recognition can be achieved when the target is missing, thereby greatly improving the robustness and accuracy of the target tracking system.

[0006] The present invention provides a target tracking method based on AI visual recognition, comprising: tracking a target in real time through satellite positioning, and when the target is missing tracking, obtaining image feature information of the target and map position information of the target when the tracking is missing, generating associated feature information based on the image feature information and the map position information; inputting the associated feature information into a preset motion prediction model, outputting the motion prediction information of the target, performing multi-size feature analysis on the associated feature information to obtain multi-scale feature information; fusing the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target, performing real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data; inputting the real-time tracking data and the missing change data into a preset AI visual recognition model to obtain real-time position information and motion trajectory information of the target.

[0007] As a preferred solution, the steps of obtaining image feature information of the target and map location information of the target when tracking is missing, and generating associated feature information based on the image feature information and the map location information, include: graying and Gaussian filtering the obtained target image to obtain a preprocessed image, and extracting features from the preprocessed image using a directional gradient histogram to obtain an initial feature set; obtaining map location information of the target to generate a location information set, performing feature matching between the initial feature set and the location information set to generate a matched feature set and an unmatched feature set; performing feature enhancement processing on the matched feature set using a SIFT algorithm to obtain an enhanced feature set; performing feature compensation on the unmatched feature set using a FAST algorithm to obtain a compensated feature set, and performing feature synthesis between the enhanced feature set and the compensated feature set to generate associated feature information.

[0008] As a preferred scheme, the step of inputting the associated feature information into a preset motion prediction model and outputting the motion prediction information of the target includes: inputting the associated feature information into a motion feature analysis module, in which the optical flow algorithm is used to extract features of the associated feature information to obtain a motion feature set and a static feature set; performing time series analysis on the motion feature set through a Kalman filter algorithm to obtain a time feature sequence, and performing spatial feature extraction on the static feature set through principal component analysis to obtain a spatial feature matrix; using a weighted fusion method to perform feature fusion on the time feature sequence and the spatial feature matrix to obtain a fused feature set, and using a Euclidean distance calculation method to perform feature matching on the fused feature set to obtain a matching feature set and matching parameters, and performing predictive calculation on the matching feature set through the matching parameters to obtain motion prediction information.

[0009] As a preferred solution, the step of performing multi-size feature analysis on the associated feature information to obtain multi-scale feature information includes: inputting the associated feature information into a multi-scale feature extraction module to obtain detail features, texture features and structural features of different scales; performing detail enhancement processing on the detail features by a Laplace pyramid detail enhancement method to obtain enhanced detail features and edge features, and performing denoising processing on the texture features by a non-local mean denoising algorithm to obtain purified texture features and shape features; fusing the enhanced detail features with the purified texture features to generate fused detail features, and combining the edge features with the shape features to generate combined shape features; applying morphological processing to process the structural features to extract morphological features and skeleton features, and performing feature fusion on the fused detail features, combined shape features, morphological features and skeleton features to obtain multi-scale feature information.

[0010] As a preferred solution, the step of fusing the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target includes: aligning the motion prediction information and the multi-scale feature information through a dynamic time warping algorithm to obtain an aligned motion feature set and an aligned multi-scale feature set; weighting the aligned motion feature set using an entropy weight method to generate weighted motion features and weight parameters; performing feature interaction between the weighted motion features and the aligned multi-scale feature set using a mutual information method to obtain an interactive feature set, and dimensionalizing the interactive feature set through a singular value decomposition algorithm to obtain an adjusted feature matrix and a residual matrix; inputting the adjusted feature matrix into a feature optimizer to extract optimized feature vectors, and weighting the optimized feature vectors using the weight parameters to generate a weighted feature matrix; fusing the weighted feature matrix with the residual matrix to obtain comprehensive feature information of the target.

[0011] As a preferred scheme, the step of performing real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data includes: tracking the target through the comprehensive feature information through a correlation filter to obtain real-time tracking data, smoothing the real-time tracking data using a mean shift algorithm to obtain smoothed tracking data; performing trajectory analysis on the smoothed tracking data according to a hidden Markov model to obtain trajectory features and a state sequence, performing missing detection on the trajectory features through a local anomaly factor algorithm to obtain missing detection results, comparing the missing detection results with the state sequence to generate missing change data.

[0012] As a preferred scheme, the step of inputting the real-time tracking data and the missing change data into a preset AI visual recognition model to obtain the real-time position information and motion trajectory information of the target includes: performing feature fusion on the real-time tracking data and the missing change data through a data fusion algorithm to obtain spatial fusion features and temporal fusion features; inputting the spatial fusion features and the temporal fusion features into a preset AI visual recognition model, and in the AI ​​visual recognition model, using a convolutional neural network to extract the spatial fusion features and the temporal fusion features to obtain a deep feature representation; using a Softmax classifier to perform probability prediction on the deep feature representation to obtain a position probability distribution and a trajectory probability distribution; applying a Bayesian inference method to perform position estimation on the position probability distribution to obtain the real-time position information of the target, and performing optimal path solution on the trajectory probability distribution according to the Viterbi algorithm to obtain the motion trajectory information of the target.

[0013] The present application also provides a target tracking device based on AI visual recognition, including: a tracking module, used to track the target in real time through satellite positioning, and when the target is missing tracking, obtain the image feature information of the target and the map position information of the target when the tracking is missing, and generate associated feature information based on the image feature information and the map position information; an analysis module, used to input the associated feature information into a preset motion prediction model, output the motion prediction information of the target, perform multi-size feature analysis on the associated feature information, and obtain multi-scale feature information; a fusion module, used to fuse the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target, perform real-time tracking and missing detection on the target based on the comprehensive feature information, and obtain real-time tracking data and missing change data; an output module, used to input the real-time tracking data and the missing change data into a preset AI visual recognition model to obtain real-time position information and motion trajectory information of the target.

[0014] The present application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements any of the above-mentioned target tracking methods based on AI visual recognition.

[0015] Compared with the prior art, the present application has the following beneficial effects: good tracking effect and strong stability. The location information of the target is obtained in real time through the satellite positioning system. When the target is lost, the feature information of the image around the target and the map location information are obtained to generate the associated feature information, and input it into the motion prediction model and multi-scale feature analysis, and output the motion prediction information and multi-scale feature information; this information is integrated to generate the comprehensive feature information of the target, and real-time tracking and missing detection are performed to obtain real-time tracking data and missing change data, and these data are analyzed by the AI ​​visual recognition model to output the real-time location information and motion trajectory information of the target; it can achieve stable tracking of the target in complex environments such as lighting changes, target occlusion or rapid movement, greatly reducing the loss of the target, improving the overall performance and application effect of tracking, and improving the problem of target loss and tracking failure when the target is in complex environments such as occlusion, lighting changes or rapid movement. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0018] Figure 1 It is a flowchart of a target tracking method based on AI visual recognition provided by an embodiment of the present invention; Figure 2 is a schematic block diagram of the structure of a target tracking device based on AI visual recognition provided by an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention.

[0019] Description of reference numerals: 10. Target tracking device based on AI visual recognition; 11. Tracking module; 12. Analysis module; 13. Fusion module; 14. Output module; 20. Electronic device; 21. Memory; 22. Processor. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are 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 creative work are within the scope of protection of the present invention.

[0021] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0022] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0023] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0024] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0025] Embodiment 1: like Figure 1 As shown, the target tracking method based on AI visual recognition provided in the embodiment of the present application includes steps S100 to S400.

[0026] Step S100: Track the target in real time through satellite positioning, and when the target is not tracked, obtain the image feature information of the target and the map position information of the target when the target is not tracked, and generate associated feature information based on the image feature information and the map position information.

[0027] In this step, the satellite positioning system is first used to obtain the precise location data of the target in real time. Specifically, when the target is detected to be lost, the camera preset on the target is automatically started to capture the environment image around the target, extract the feature information in the image, and record the map location information of the target when it is lost. Based on the extracted feature information and map location information, the associated feature information is generated for subsequent motion prediction and tracking operations.

[0028] For example, when the target loses satellite signals while driving in complex terrain or in a tunnel, the camera will capture characteristic information of the interior of the tunnel and the entrances and exits, and generate associated characteristic information based on the last recorded target location information.

[0029] Step S200: input the associated feature information into a preset motion prediction model, output the motion prediction information of the target, perform multi-scale feature analysis on the associated feature information, and obtain multi-scale feature information.

[0030] In this step, the generated associated feature information is input into the preset motion prediction model, and the motion prediction information of the target is obtained through model calculation. Specifically, the model will perform motion trend analysis based on the associated feature information and predict the motion trajectory of the target. At the same time, multi-scale feature analysis is performed on the associated feature information to extract key information at different scales and obtain multi-scale feature information, providing data support for the generation of comprehensive feature information.

[0031] For example, the motion prediction model can predict the target's direction of travel and speed changes within a certain period of time, and obtain detailed feature information of the target's surrounding environment, such as building outlines and landmark locations, through multi-scale feature analysis.

[0032] Step S300: Fusing the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target, and performing real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data.

[0033] In this step, the comprehensive feature information of the target is generated by fusing the motion prediction information and the multi-scale feature information. Specifically, the fused comprehensive feature information is used to continuously track and detect missing objects in real time to ensure the accuracy of the positioning and tracking of the target in complex environments. At the same time, real-time tracking data and missing change data are generated to provide basic data for subsequent visual recognition.

[0034] For example, the fused comprehensive feature information can reflect the characteristic changes of the target under various environmental factors. For example, when the lighting changes or is blocked, the position of the target can still be accurately identified and tracked.

[0035] Step S400: input the real-time tracking data and missing change data into a preset AI visual recognition model to obtain the real-time location information and motion trajectory information of the target.

[0036] In this step, the real-time tracking data and missing change data are input into the preset AI visual recognition model, and the recognition ability of the model is used to accurately locate the target. Specifically, the model analyzes the input data and outputs the real-time location information and motion trajectory information of the target to achieve continuous monitoring of the target.

[0037] For example, the AI ​​visual recognition model can identify the current location of the target based on real-time tracking data and predict the target's next movement direction through missing change data, ensuring continuous and effective monitoring of the target.

[0038] In this embodiment, the target is tracked in real time by satellite positioning. When the target is lost, the camera preset on the target is started to obtain the feature information of the target's surrounding image and the map location information of the target when the tracking is lost, and then the associated feature information is generated based on these feature information and map location information. Next, the associated feature information is input into the preset motion prediction model, the motion prediction information of the target is output, and the associated feature information is subjected to multi-scale feature analysis to obtain multi-scale feature information. Finally, the motion prediction information and the multi-scale feature information are fused to obtain the comprehensive feature information of the target, and the target is tracked and missing detection is performed on the target in real time based on the comprehensive feature information, so as to obtain real-time tracking data and missing change data, and input them into the preset AI visual recognition model to obtain the real-time location information and motion trajectory information of the target. By timely remediation when the target is lost, the accuracy and reliability of target tracking are improved. Combining a variety of technical means, it is ensured that the target can still be stably tracked in complex environments, such as in the case of illumination changes, occlusion or rapid movement, through the application of multi-scale feature analysis and motion prediction models, the situation of target loss is greatly reduced, and the overall performance and application effect of tracking are improved.

[0039] The location information of the target is obtained in real time through the satellite positioning system. When the target is lost, the camera is started to obtain the feature information and map location information of the target's surrounding images, generate associated feature information, and input it into the motion prediction model and multi-scale feature analysis, and output motion prediction information and multi-scale feature information. After that, this information is fused to generate comprehensive feature information of the target, and real-time tracking and missing detection are performed to obtain real-time tracking data and missing change data. Finally, these data are analyzed by the AI ​​visual recognition model to output the real-time location information and motion trajectory information of the target. This method can maintain stable tracking of the target in complex environments such as lighting changes, target occlusion or rapid movement, greatly reducing the loss of the target, improving the overall performance and application effect of tracking, and improving the problem of target loss and tracking failure when the target is in complex environments such as occlusion, lighting changes or rapid movement.

[0040] Embodiment 2: In step S100, the image feature information of the target and the map position information of the target when tracking is missing are obtained, and the step of generating associated feature information based on the image feature information and the map position information includes: graying and Gaussian filtering the obtained target image to obtain a preprocessed image, and using a directional gradient histogram to extract features from the preprocessed image to obtain an initial feature set.

[0041] By graying the acquired target image, color interference is eliminated and the efficiency of image processing is improved; specifically, the graying technology is used to convert the color image into a gray image, thereby reducing the computational complexity. Then, the Gaussian filtering method is applied to smooth the gray image to reduce the noise in the image and obtain the preprocessed image. Subsequently, the directional gradient histogram is used to extract features from the preprocessed image to capture the edge and texture information in the image and form an initial feature set.

[0042] For example, in image preprocessing, the vehicle image captured by the camera is grayed and Gaussian filtered to extract initial features such as contour and texture.

[0043] The map location information of the target is obtained, a location information set is generated, and the initial feature set is matched with the location information set to generate a matched feature set and an unmatched feature set.

[0044] By using GPS and map data to obtain the precise location information of the target, a location information set containing the target's geographical location is generated; specifically, the target location is converted into a location information set by combining satellite positioning data and real-time maps. Then, the image features in the initial feature set are matched with the location information set, and a matching feature set and an unmatched feature set are generated through a feature matching algorithm.

[0045] For example, in vehicle tracking, the image features of the vehicle and its location information on the driving path are combined to match and generate a set of matching features that include the current state of the vehicle.

[0046] The matching feature set is enhanced by using the SIFT algorithm to obtain an enhanced feature set.

[0047] The SIFT (Scale Invariant Feature Transform) algorithm is used to enhance the matching feature set to improve the stability and robustness of the feature points. Specifically, the SIFT algorithm is used to extract key points in the image and calculate the descriptors of these key points, so that the feature points have higher stability in scale, rotation and illumination changes, and an enhanced feature set is generated.

[0048] For example, during the enhancement process, the SIFT algorithm is used to extract feature points in the vehicle image and calculate its descriptor, thereby enhancing the image features of the vehicle.

[0049] The unmatched feature set is compensated by the FAST algorithm to obtain a compensated feature set, and the enhanced feature set and the compensated feature set are synthesized to generate associated feature information.

[0050] By applying the FAST (Features from Accelerated Segment Test) algorithm to the unmatched feature set for feature compensation, more key feature points can be identified; specifically, the FAST algorithm is used to quickly detect corner points in the image and perform feature compensation to generate a compensated feature set. Subsequently, the enhanced feature set and the compensated feature set are synthesized to generate more comprehensive associated feature information.

[0051] For example, when processing unmatched features, the FAST algorithm can quickly detect unmatched corner features in vehicle images and synthesize them with the enhanced feature set to form complete associated feature information.

[0052] In step S200, the associated feature information is input into a preset motion prediction model, and the step of outputting the motion prediction information of the target includes: inputting the associated feature information into a motion feature analysis module, and in the motion feature analysis module, using an optical flow algorithm to extract features from the associated feature information to obtain a motion feature set and a static feature set.

[0053] The generated associated feature information is input into the motion feature analysis module, and the optical flow algorithm is used to extract its features and analyze the motion information of the target. Specifically, the optical flow algorithm calculates the motion vector of the pixels in the image sequence, extracts the motion feature set and the static feature set, thereby distinguishing the moving part and the stationary part of the target.

[0054] For example, in vehicle tracking, the optical flow algorithm can be used to extract the vehicle's motion trajectory and the surrounding static environmental features.

[0055] The Kalman filter algorithm is used to perform time series analysis on the motion feature set to obtain the time feature sequence, and the principal component analysis is used to extract spatial features from the static feature set to obtain the spatial feature matrix.

[0056] The motion trajectory and state of the target are predicted by applying the Kalman filter algorithm to the motion feature set for time series analysis. Specifically, the Kalman filter iteratively calculates the time series data to obtain the time feature sequence of the target. Subsequently, the principal component analysis (PCA) is used to reduce the dimension of the static feature set, extract the main features, and form a spatial feature matrix.

[0057] For example, Kalman filtering can be used to predict the position of a vehicle at a future time, while the PCA method can extract the main features of the vehicle's surroundings.

[0058] The weighted fusion method is used to fuse the time feature sequence and the spatial feature matrix to obtain a fused feature set. The Euclidean distance calculation method is used to perform feature matching on the fused feature set to obtain a matching feature set and matching parameters. The matching feature set is predicted and calculated using the matching parameters to obtain motion prediction information.

[0059] By adopting the weighted fusion method, the temporal feature sequence and the spatial feature matrix are comprehensively analyzed to improve the accuracy of the prediction; specifically, the weighted fusion method is used to assign weights according to the importance of the features, and the temporal and spatial features are fused to generate a fused feature set. Then, the Euclidean distance calculation method is used to match the fused feature set, determine the motion state of the target, and obtain the matching feature set and matching parameters. The matching feature set is predicted and calculated through the matching parameters to generate motion prediction information.

[0060] For example, in the motion prediction process, the weighted fusion method can more accurately predict the vehicle's trajectory, and the Euclidean distance matching method is used to confirm the vehicle's current position.

[0061] In step S200, the step of performing multi-size feature analysis on the associated feature information to obtain multi-scale feature information includes: inputting the associated feature information into a multi-scale feature extraction module to obtain detail features, texture features and structural features of different scales.

[0062] By inputting the associated feature information into the multi-scale feature extraction module, features of different scales are extracted from the image; specifically, through the multi-scale feature extraction module, detail features, texture features and structural features in the image are extracted respectively to form complete multi-scale feature information.

[0063] For example, the multi-scale feature extraction module can extract the vehicle's detailed texture, surface texture, and overall contour structure from the image.

[0064] The detail features are enhanced by the Laplacian pyramid detail enhancement method to obtain enhanced detail features and edge features. The texture features are denoised by the non-local mean denoising algorithm to obtain purified texture features and shape features.

[0065] By applying the Laplacian pyramid detail enhancement method, the detail features are enhanced to improve the clarity of the features; specifically, the Laplacian pyramid method enhances the details and edge features in the image by decomposing and reconstructing the image, generating enhanced detail features. Then, the texture features are denoised using the non-local mean denoising algorithm to eliminate image noise and generate purified texture features and shape features.

[0066] For example, the Laplacian pyramid method can be used to enhance the details in the vehicle image and make the license plate number clearer, and the non-local means denoising algorithm is used to remove noise in the image and improve the quality of texture features.

[0067] The enhanced detail features are fused with the purified texture features to generate fused detail features, and the edge features are combined with the shape features to generate combined shape features.

[0068] By fusing the enhanced detail features with the purified texture features, a more complete fused detail feature is generated; specifically, through feature fusion technology, the enhanced detail part of the image is combined with the purified texture part to form a clearer feature set. Then, the edge features are combined with the shape features to generate a combined shape feature, further improving the overall feature representation of the image.

[0069] For example, after being fused, the detail features and texture features in a vehicle image can better reflect the overall appearance of the vehicle, while the combination of edge features and shape features can accurately describe the outline and shape of the vehicle.

[0070] Morphological processing is applied to process the structural features, morphological features and skeleton features are extracted, and the fusion of detail features, combined shape features, morphological features and skeleton features is performed to obtain multi-scale feature information.

[0071] By applying morphological processing technology, the structural features are processed to extract the morphological features and skeleton features of the image; specifically, morphological processing performs structural analysis on the image through operations such as expansion, corrosion, opening and closing operations to extract morphological features and skeleton features. Enhanced detail features, combined shape features, morphological features and skeleton features are fused to generate complete multi-scale feature information.

[0072] For example, morphological processing can extract the morphological features and skeleton features of a vehicle, and fuse them with detail features and shape features to generate more comprehensive multi-scale feature information of the vehicle.

[0073] In step S300, the step of fusing the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target includes: aligning the motion prediction information and the multi-scale feature information through a dynamic time warping algorithm to obtain an aligned motion feature set and an aligned multi-scale feature set.

[0074] The dynamic time warping (DTW) algorithm is used to align motion prediction information and multi-scale feature information to solve the time offset problem between different data sources. Specifically, the DTW algorithm adjusts the matching path of the time series through dynamic programming technology to align feature data of different scales and obtain the aligned motion feature set and multi-scale feature set.

[0075] For example, in vehicle tracking, the DTW algorithm can align the vehicle's motion trajectory information and multi-scale image features to ensure data consistency.

[0076] The entropy weight method is used to weight the aligned motion feature set to generate weighted motion features and weight parameters.

[0077] By adopting the entropy weight method, the aligned motion feature set is weighted to determine the importance of each feature; specifically, the entropy weight method assigns weights according to the information entropy value of each feature, thereby generating weighted motion features and weight parameters to enhance the representativeness of the features.

[0078] For example, the entropy weight method can assign appropriate weights according to the changes in the vehicle's motion characteristics, so that important motion features have a greater influence in subsequent processing.

[0079] The mutual information method is used to perform feature interaction between the weighted motion features and the aligned multi-scale feature set to obtain an interactive feature set. The singular value decomposition algorithm is used to adjust the dimension of the interactive feature set to obtain the adjusted feature matrix and residual matrix.

[0080] The mutual information method is used to perform feature interaction between weighted motion features and aligned multi-scale feature sets to enhance the correlation between features. Specifically, the mutual information method calculates the mutual information between features, determines the correlation between features, and generates an interactive feature set. Then, the singular value decomposition (SVD) algorithm is used to adjust the dimension of the interactive feature set to reduce feature redundancy and obtain the adjusted feature matrix and residual matrix.

[0081] For example, the mutual information method can enhance the association between vehicle motion features and image features, and the SVD algorithm is used to simplify feature data and improve processing efficiency.

[0082] The adjusted feature matrix is ​​input into the feature optimizer, the optimized feature vector is extracted, and the optimized feature vector is weighted using the weight parameter to generate a weighted feature matrix.

[0083] By inputting the adjusted feature matrix into the feature optimizer, the feature vector is optimized to improve the recognition of the feature; specifically, the feature optimizer extracts the optimized feature vector through the optimization algorithm to reduce noise interference. Subsequently, the optimized feature vector is weighted using the weight parameter to generate a weighted feature matrix to further improve the representativeness of the feature.

[0084] For example, the feature optimizer can extract the most distinctive features in a vehicle image and enhance the influence of these features through weighted processing.

[0085] The weighted feature matrix is ​​fused with the residual matrix to obtain the comprehensive feature information of the target.

[0086] By fusing the weighted feature matrix with the residual matrix, more complete comprehensive feature information is formed; specifically, through the feature fusion technology, the weighted feature matrix is ​​combined with the residual matrix to comprehensively reflect the motion and multi-scale characteristics of the target and generate the comprehensive feature information of the target.

[0087] For example, in vehicle tracking, the fused comprehensive feature information can simultaneously reflect the vehicle's motion trajectory and image features, improving tracking accuracy.

[0088] In step S300, the step of performing real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data includes: tracking the target through the comprehensive feature information through a correlation filter to obtain real-time tracking data, and smoothing the real-time tracking data using a mean shift algorithm to obtain smoothed tracking data.

[0089] The correlation filter is used to track the target based on the comprehensive feature information to ensure the accuracy of real-time tracking. Specifically, the correlation filter determines the position of the target by calculating the correlation between the comprehensive feature information and the target template, and generates real-time tracking data. Then, the mean shift algorithm is used to smooth the real-time tracking data to reduce data fluctuations and obtain smooth tracking data.

[0090] For example, the correlation filter can accurately track the position changes of the vehicle, and the mean shift algorithm is used to smooth the vehicle's motion trajectory and reduce noise interference.

[0091] The smooth tracking data is analyzed by trajectory analysis according to the hidden Markov model to obtain trajectory features and state sequences. The trajectory features are detected for missingness through the local anomaly factor algorithm to obtain missingness detection results. The missingness detection results are compared with the state sequences to generate missingness change data.

[0092] By applying the Hidden Markov Model (HMM) to perform trajectory analysis on smooth tracking data, the target's motion pattern can be identified; specifically, HMM generates trajectory features and state sequences by analyzing the state transition of smooth data. Then, the local outlier factor (LOF) algorithm is used to perform missing detection on the trajectory features, identify abnormal behaviors, and generate missing detection results. Finally, the missing detection results are compared with the state sequence to generate missing change data reflecting the target's motion changes.

[0093] For example, HMM can analyze the vehicle's driving path, and the LOF algorithm is used to detect abnormal stops or sudden accelerations in vehicle movement and generate missing change data.

[0094] In step S400, the real-time tracking data and missing change data are input into a preset AI visual recognition model to obtain the real-time position information and motion trajectory information of the target, including: performing feature fusion on the real-time tracking data and missing change data through a data fusion algorithm to obtain spatial fusion features and temporal fusion features.

[0095] By adopting the data fusion algorithm, a comprehensive analysis is performed on the real-time tracking data and the missing change data; specifically, the data fusion algorithm fuses the real-time tracking data and the missing change data by combining spatial and temporal features to generate fused features containing spatial and temporal information.

[0096] For example, the data fusion algorithm can integrate the vehicle's real-time position and trajectory change data to generate fusion features that reflect the vehicle's motion state.

[0097] The spatial fusion features and the temporal fusion features are input into a preset AI visual recognition model. In the AI ​​visual recognition model, a convolutional neural network is used to extract the spatial fusion features and the temporal fusion features to obtain a deep feature representation.

[0098] Deep learning analysis is performed by inputting spatial fusion features and temporal fusion features into a preset AI visual recognition model; specifically, the AI ​​visual recognition model uses a convolutional neural network (CNN) to extract fusion features to obtain a higher-level deep feature representation.

[0099] For example, CNN can extract complex features and motion information in vehicle images and generate deep feature representations for subsequent recognition and analysis.

[0100] The Softmax classifier is used to perform probability prediction on the deep feature representation to obtain the position probability distribution and trajectory probability distribution.

[0101] The probability distribution of the target position and trajectory is determined by using the Softmax classifier to make probability predictions on the deep feature representation. Specifically, the Softmax classifier calculates the probability of each possible position and trajectory based on the input deep feature representation to generate the position probability distribution and trajectory probability distribution. These probability distributions reflect the possibility of the target at different positions and paths.

[0102] For example, in vehicle tracking, a Softmax classifier can predict the probability distribution of a vehicle at different locations, as well as its possible trajectory, based on the deep feature representation.

[0103] The Bayesian inference method is used to estimate the position probability distribution and obtain the real-time position information of the target. The optimal path of the trajectory probability distribution is solved according to the Viterbi algorithm to obtain the motion trajectory information of the target.

[0104] The Bayesian inference method is used to analyze the position probability distribution and estimate the real-time position of the target. Specifically, the Bayesian inference method calculates the most likely position of the target based on the position probability distribution and obtains the real-time position information of the target. Then, the Viterbi algorithm is used to solve the optimal path for the trajectory probability distribution, identify the target's motion trajectory, and ensure the accuracy and consistency of the path.

[0105] For example, the real-time position of a vehicle can be accurately located through the Bayesian reasoning method, and the Viterbi algorithm is used to identify the optimal driving path of the vehicle and generate continuous motion trajectory information.

[0106] In this embodiment, by combining satellite positioning, camera acquisition of image feature information, map location information, motion prediction model, multi-scale feature analysis, and AI visual recognition model and other technical means, timely remediation can be performed when the target is lost, thereby improving the accuracy and reliability of target tracking. In the specific implementation, the location information of the target is obtained in real time through the satellite positioning system. When the target is lost, the camera is started to obtain the feature information and map location information of the image around the target, generate associated feature information, and input it into the motion prediction model and multi-scale feature analysis, and output motion prediction information and multi-scale feature information. After that, this information is fused to generate comprehensive feature information of the target, and real-time tracking and missing detection are performed to obtain real-time tracking data and missing change data. Finally, these data are analyzed by the AI ​​visual recognition model to output the real-time location information and motion trajectory information of the target. This method can still maintain stable tracking of the target in complex environments such as lighting changes, target obstruction or rapid movement, greatly reducing the situation of target loss and improving the overall performance and application effect of tracking.

[0107] Embodiment 3: like Figure 2 As shown, the present application also provides a target tracking device 10 based on AI visual recognition, including a tracking module 11, an analysis module 12, a fusion module 13 and an output module 14.

[0108] The tracking module 11 is mainly used to track the target in real time through satellite positioning, and when the target is not tracked, obtain the image feature information of the target and the map position information of the target when the target is not tracked, and generate associated feature information based on the image feature information and the map position information.

[0109] The analysis module 12 is mainly used to input the associated feature information into a preset motion prediction model, output the motion prediction information of the target, perform multi-scale feature analysis on the associated feature information, and obtain multi-scale feature information.

[0110] The fusion module 13 is mainly used to fuse the motion prediction information and the multi-scale feature information to obtain the comprehensive feature information of the target, and to perform real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data.

[0111] The output module 14 is mainly used to input the real-time tracking data and missing change data into a preset AI visual recognition model to obtain the real-time location information and motion trajectory information of the target.

[0112] In this embodiment, high-precision real-time tracking of the target is achieved by combining the coordinated operation of multiple functional modules such as the tracking module 11, the analysis module 12, the fusion module 13 and the output module 14. When the target is lost, the feature information and map location information are obtained through the camera, and the associated feature information is generated. The feature information is analyzed in multiple sizes and scales through the analysis module 12, and the motion prediction information and multi-scale feature information of the target are output. The fusion module 13 comprehensively processes this information, generates comprehensive feature information of the target, and performs real-time tracking and missing detection on the target. Finally, the output module 14 inputs the real-time tracking data and missing change data into the AI ​​visual recognition model to obtain the real-time location information and motion trajectory information of the target. The present device can efficiently and accurately track targets in complex environments, improving the overall performance and reliability of target tracking.

[0113] It should be noted that technical personnel in the relevant technical field can clearly understand that, for the convenience and conciseness of description, the specific working process of the above-described device and each module can refer to the corresponding process in the aforementioned target tracking method based on AI visual recognition embodiment, and will not be repeated here.

[0114] Embodiment 4: like Figure 3 As shown, the present application also provides an electronic device 20, including a memory 21 and a processor 22, wherein the memory 21 stores a computer program that can be run on the processor 22, and when the processor 22 executes the computer program, the target tracking method based on AI visual recognition of Example 1 is implemented.

[0115] In this embodiment, a target tracking method based on AI visual recognition is implemented by combining the memory 21 and the processor 22. Specifically, the computer program stored in the memory 21 includes an instruction set for performing satellite positioning, image processing, motion prediction, multi-scale feature analysis, feature fusion and AI visual recognition. When the processor 22 executes these instructions, the target can be accurately tracked in a real-time environment. When the target is lost, the device will automatically start the camera to obtain image feature information and map location information, generate comprehensive feature information through motion prediction model and multi-scale feature analysis, and perform real-time tracking and missing detection of the target. Finally, the real-time position information and motion trajectory information of the target are calculated and output through the AI ​​visual recognition model. The electronic device 20 can efficiently track targets in complex and dynamic environments, greatly improving the accuracy and stability of target tracking.

[0116] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target tracking method based on AI visual recognition, characterized in that: include: Tracking a target in real time through satellite positioning, and when the target is not tracked, obtaining image feature information of the target and map location information of the target when tracking is not available, and generating associated feature information based on the image feature information and the map location information; Inputting the associated feature information into a preset motion prediction model, outputting motion prediction information of the target, and performing multi-scale feature analysis on the associated feature information to obtain multi-scale feature information; The motion prediction information and the multi-scale feature information are integrated to obtain comprehensive feature information of the target, and the target is tracked and missing information is detected in real time based on the comprehensive feature information to obtain real-time tracking data and missing change data; The real-time tracking data and the missing change data are input into a preset AI visual recognition model to obtain the real-time location information and motion trajectory information of the target.

2. The target tracking method based on AI visual recognition according to claim 1, characterized in that: The step of acquiring the image feature information of the target and the map position information of the target when tracking is missing, and generating associated feature information based on the image feature information and the map position information comprises: The acquired target image is grayed and preprocessed by Gaussian filtering to obtain a preprocessed image, and a directional gradient histogram is used to extract features of the preprocessed image to obtain an initial feature set; Acquire map location information of the target, generate a location information set, perform feature matching between the initial feature set and the location information set, and generate a matching feature set and an unmatched feature set; Performing feature enhancement processing on the matching feature set by using a SIFT algorithm to obtain an enhanced feature set; The unmatched feature set is feature compensated by using the FAST algorithm to obtain a compensated feature set, and the enhanced feature set and the compensated feature set are feature synthesized to generate associated feature information.

3. The target tracking method based on AI visual recognition according to claim 1, characterized in that: The step of inputting the associated feature information into a preset motion prediction model and outputting the motion prediction information of the target comprises: The associated feature information is input into a motion feature analysis module, and in the motion feature analysis module, an optical flow algorithm is used to extract features from the associated feature information to obtain a motion feature set and a static feature set; Performing time series analysis on the motion feature set by using a Kalman filter algorithm to obtain a time feature sequence, and performing spatial feature extraction on the static feature set by using principal component analysis to obtain a spatial feature matrix; A weighted fusion method is used to fuse the time feature sequence and the spatial feature matrix to obtain a fused feature set. The Euclidean distance calculation method is used to perform feature matching on the fused feature set to obtain a matching feature set and matching parameters. The matching feature set is predicted and calculated using the matching parameters to obtain motion prediction information.

4. The target tracking method based on AI visual recognition according to claim 1, characterized in that: The step of performing multi-scale feature analysis on the associated feature information to obtain multi-scale feature information includes: Inputting the associated feature information into a multi-scale feature extraction module to obtain detail features, texture features and structural features at different scales; Performing detail enhancement processing on the detail features by using a Laplace pyramid detail enhancement method to obtain enhanced detail features and edge features, and performing denoising processing on the texture features by using a non-local mean denoising algorithm to obtain purified texture features and shape features; Fusing the enhanced detail feature with the purified texture feature to generate a fused detail feature, and combining the edge feature with the shape feature to generate a combined shape feature; The structural features are processed by applying morphological processing, morphological features and skeleton features are extracted, and the fused detail features, combined shape features, morphological features and skeleton features are fused to obtain multi-scale feature information.

5. The target tracking method based on AI visual recognition according to claim 1, characterized in that: The step of fusing the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target includes: Performing feature alignment on the motion prediction information and the multi-scale feature information by a dynamic time warping algorithm to obtain an aligned motion feature set and an aligned multi-scale feature set; Performing weighted processing on the aligned motion feature set using an entropy weight method to generate weighted motion features and weight parameters; Using a mutual information method to perform feature interaction between the weighted motion features and the aligned multi-scale feature set to obtain an interactive feature set, and using a singular value decomposition algorithm to perform dimension adjustment on the interactive feature set to obtain an adjusted feature matrix and a residual matrix; Inputting the adjusted feature matrix into a feature optimizer, extracting optimized feature vectors, and performing weighted processing on the optimized feature vectors using the weight parameters to generate a weighted feature matrix; The weighted feature matrix is ​​fused with the residual matrix to obtain comprehensive feature information of the target.

6. The target tracking method based on AI visual recognition according to claim 1, characterized in that: The step of performing real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data includes: Tracking the comprehensive feature information by using a correlation filter to obtain real-time tracking data, and smoothing the real-time tracking data by using a mean shift algorithm to obtain smoothed tracking data; The smoothed tracking data is subjected to trajectory analysis according to a hidden Markov model to obtain trajectory features and a state sequence, the trajectory features are subjected to missing detection through a local anomaly factor algorithm to obtain missing detection results, the missing detection results are compared with the state sequence to generate missing change data.

7. The target tracking method based on AI visual recognition according to claim 1, characterized in that: The step of inputting the real-time tracking data and the missing change data into a preset AI visual recognition model to obtain the real-time location information and motion trajectory information of the target includes: Performing feature fusion on the real-time tracking data and the missing change data through a data fusion algorithm to obtain spatial fusion features and temporal fusion features; Inputting the spatial fusion features and the temporal fusion features into a preset AI visual recognition model, and in the AI ​​visual recognition model, using a convolutional neural network to perform feature extraction on the spatial fusion features and the temporal fusion features to obtain a deep feature representation; A Softmax classifier is used to perform probability prediction on the deep feature representation to obtain a position probability distribution and a trajectory probability distribution; The Bayesian inference method is used to estimate the position probability distribution to obtain the real-time position information of the target, and the optimal path of the trajectory probability distribution is solved according to the Viterbi algorithm to obtain the motion trajectory information of the target.

8. A target tracking device based on AI visual recognition, characterized in that: include: A tracking module, used to track a target in real time through satellite positioning, and when the target is not tracked, obtain image feature information of the target and map location information of the target when tracking is not available, and generate associated feature information based on the image feature information and the map location information; An analysis module, used to input the associated feature information into a preset motion prediction model, output the motion prediction information of the target, and perform multi-scale feature analysis on the associated feature information to obtain multi-scale feature information; A fusion module, used to fuse the motion prediction information and the multi-scale feature information to obtain comprehensive feature information of the target, and perform real-time tracking and missing detection on the target based on the comprehensive feature information to obtain real-time tracking data and missing change data; The output module is used to input the real-time tracking data and the missing change data into a preset AI visual recognition model to obtain the real-time location information and motion trajectory information of the target.

9. An electronic device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the target tracking method based on AI visual recognition as described in any one of claims 1 to 7 when executing the computer program.

Images