An intelligent park target identification method and system based on artificial intelligence

Abstract

This invention belongs to the field of artificial intelligence technology, specifically relating to an AI-based smart park target recognition method and system. The method includes the following steps: accessing the park's surveillance video stream, analyzing the camera topology to determine blind zone channels, obtaining the physical path length and illumination change intensity index of the blind zone channels through map mapping and historical image analysis, and simultaneously extracting the target's trajectory point sequence within the visible area; monitoring the target's dwell time after entering the blind zone, and calculating the confidence weight of visual features based on the target's speed fluctuations before entering the blind zone and the environmental complexity factors of the blind zone channel, using a preset feature effectiveness attenuation model. The confidence weight decreases non-linearly with increasing dwell time. This invention effectively solves the problems of appearance failure and large motion prediction deviations caused by long dwell times, significantly improving recognition accuracy.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a smart park target recognition method and system based on artificial intelligence. Background Technology

[0002] In the security monitoring and management system of smart parks, continuous tracking and target re-identification of moving targets such as pedestrians and vehicles are core requirements for ensuring park safety and operational efficiency. Existing target recognition systems mainly rely on cameras distributed throughout the park to capture target images, extract appearance features for comparison, and achieve cross-camera tracking of targets.

[0003] However, the complex environment of a park inevitably creates blind spots in the camera network, such as obstructed areas behind buildings, tree-lined paths, or unmonitored connecting passages. Existing technologies typically employ fixed weighting strategies to fuse visual feature similarity and spatiotemporal matching when processing target association across blind spots. However, the longer a target lingers in a blind spot, the higher the probability of it changing posture, being obscured by clothing, or being affected by ambient lighting. In this case, the reliability of the visual features drops sharply. Existing technologies ignore this dynamic decay, leading to misjudgments in long-staying scenarios due to over-reliance on visual features.

[0004] Furthermore, existing technologies for predicting motion across blind zones often simplify by assuming that the target moves at a constant velocity in a straight line within the blind zone. In reality, the uncertainty of the target's motion increases non-linearly with increasing dwell time. Simple linear prediction models cannot accommodate the positional deviations caused by long dwell times, leading to incorrect target filtering and failing to meet the requirements of high-precision security monitoring. Summary of the Invention

[0005] This invention provides a smart park target recognition method and system based on artificial intelligence to solve the technical problem that existing target recognition technologies suffer from low accuracy in cross-blind zone target association due to the failure of appearance features and large deviations in motion prediction caused by long-term dwell time in park blind zone scenarios.

[0006] In a first aspect, the present invention provides a smart park target recognition method based on artificial intelligence, comprising the following steps:

[0007] Access the park's surveillance video stream, analyze the camera topology to determine blind spot channels, and obtain the physical path length and light intensity index of blind spot channels through map mapping and historical image analysis. At the same time, extract the trajectory point sequence of the target within the visible area.

[0008] The system monitors the dwell time of a target after it enters the blind zone. Based on the speed fluctuations of the target before entering the blind zone and the environmental complexity factors of the blind zone channel, the system calculates the confidence weight of visual features using a preset feature effectiveness attenuation model. The confidence weight decreases non-linearly with the increase of dwell time.

[0009] Based on the instantaneous velocity of the target when it enters the blind zone and the physical path length of the blind zone channel, a probabilistic prediction model including a time spread term is used to calculate the spatiotemporal arrival probability of the target reaching the next camera after the dwell time. The time spread term is used to evaluate the range of prediction error as time increases.

[0010] When the next camera captures a candidate target, the visual similarity between the candidate target and the original target is calculated. The visual similarity and the spatiotemporal arrival probability are dynamically weighted and fused using confidence weights to obtain the final matching score, thereby determining the target's identity.

[0011] Its effects are as follows: By introducing a mechanism that allows the confidence level of visual features to decay non-linearly with dwell time, this invention solves the problem of decreased recognition rate caused by appearance changes after a target has been in the blind zone for a long time in existing technologies; at the same time, by using a probabilistic prediction model that includes a time diffusion term, it makes up for the deficiency of traditional linear prediction in not being able to accommodate non-uniform motion deviations; by dynamically fusing visual similarity and spatiotemporal arrival probability according to confidence level, it significantly improves the accuracy of cross-blind zone target association in the complex environment of smart parks.

[0012] Furthermore, the physical path length and illumination change intensity index of the blind zone channel are obtained, including:

[0013] The non-visual area between two adjacent cameras in the camera topology is defined as the blind zone channel.

[0014] The physical path length of the blind spot passage is obtained through map surveying;

[0015] By comparing the historical average brightness histograms of the exit of the upper camera and the entrance of the lower camera in two adjacent cameras, and normalizing the comparison results, the intensity index of light change is obtained.

[0016] Its effect is as follows: By mapping the physical path of the blind zone and normalizing the difference in illumination between adjacent cameras, the present invention achieves accurate quantification of the physical properties of the blind zone environment. This not only provides an accurate distance benchmark for subsequent motion prediction, but also effectively assesses the degree of damage to visual features caused by sudden changes in illumination, and provides objective and reliable data support for the calculation of complex environmental factors, thereby enhancing the system's ability to perceive environmental differences.

[0017] Furthermore, the confidence weights of visual features are calculated, including:

[0018] Extract the instantaneous velocity sequence of the target within a preset number of frames before entering the blind zone, calculate the standard deviation of the instantaneous velocity sequence, and obtain the velocity fluctuation value;

[0019] The environmental complexity factor is calculated by combining the number of branch roads in the blind zone and the intensity index of light change.

[0020] By substituting the speed fluctuation value, environmental complexity factor, and residence time into the feature effectiveness decay model, the confidence weight is calculated.

[0021] Its effect is that the present invention constructs a multi-dimensional environmental complexity factor by comprehensively considering the speed fluctuation of the target before entering the blind zone, as well as the number of branch paths and changes in illumination within the blind zone. This makes the confidence assessment of visual features no longer solely dependent on time, but can adaptively adjust according to the severity of the specific scenario and the uncertainty of the target's behavior, thereby more accurately reflecting the true usability of visual features in the current specific blind zone environment.

[0022] Furthermore, the feature effectiveness decay model is as follows:

[0023]

[0024] In the formula, The confidence weights for visual features are... Based on the stability constant, As a complex environmental factor, For the duration of stay, The velocity fluctuation value before the target enters. It is the logarithmic safety constant.

[0025] Its effects are as follows: This invention uses a specific mathematical model to quantify the nonlinear relationship between visual confidence, dwell time, and environmental factors. By introducing a logarithmic function and a smoothing constant, it ensures the numerical stability of the weight values ​​under different orders of magnitude of input. The model can accurately simulate the physical process of target features gradually becoming ineffective with time and environmental interference, providing a scientific and smooth quantitative indicator for the subsequent weighted fusion of visual and spatiotemporal features.

[0026] Furthermore, the spatiotemporal arrival probability of the target reaching the next camera after the dwell time is calculated, including:

[0027] Obtain the instantaneous velocity of the target when it enters the blind zone;

[0028] Calculate the absolute deviation between the theoretical travel distance and the physical path length of the target during the dwell time;

[0029] The standard deviation of the path length is corrected using the time diffusion term to obtain the dynamic error range;

[0030] The spatiotemporal arrival probability is calculated based on the ratio of absolute deviation to dynamic error range.

[0031] Its effects are as follows: This invention introduces a time diffusion mechanism to dynamically correct the standard deviation of the physical path length, acknowledging and quantifying the prediction error range that inevitably expands with the increase of dwell time; this approach gives the prediction model the fault tolerance capability that conforms to physical laws, avoids the system from incorrectly filtering out arrival time deviations caused by non-uniform target movement or midway stops, and significantly improves the ability to capture long-staying or atypical moving targets.

[0032] Furthermore, the formula for calculating the spatiotemporal arrival probability is:

[0033]

[0034] In the formula, For the probability of arrival in spacetime, The instantaneous velocity of the target when it enters the blind spot. The physical length of the blind spot passage. For the duration of stay, The standard deviation of the path length is the baseline value. is the time diffusion coefficient.

[0035] Its effect is as follows: This invention maps the spatiotemporal rationality of the target reaching the next camera to a probability value between 0 and 1 through an exponential probability density function; this formula accurately describes the degree of deviation between the actual motion and the theoretical motion of the target. Combined with the time diffusion coefficient, it can maintain high-precision constraints during short-term dwell time and automatically relax the judgment boundary during long-term dwell time, realizing dynamic relaxation of spatiotemporal constraints, which is more in line with the actual physical motion law.

[0036] Furthermore, the formula for calculating the final matching score is as follows:

[0037]

[0038] In the formula, For the final match score, The confidence weights for visual features are... The visual similarity between the candidate target and the original target. For the probability of arrival in spacetime, This is the dimensional balance coefficient.

[0039] Furthermore, the calculation process for environmental complexity factors includes:

[0040] Obtain the number of branch paths within the blind zone and the maximum baseline number of branch paths within the blind zone;

[0041] Calculate the ratio of the number of branch roads to the maximum base number of branch roads;

[0042] By assigning preset weights to the light intensity change index and ratio and summing them, the environmental complexity factor is obtained.

[0043] Furthermore, the calculation process for the velocity fluctuation value includes:

[0044] Acquire the last preset number of video frames before the target enters the blind zone;

[0045] The coordinates of the center point of the target in each frame are extracted using a target detection algorithm. The ratio of displacement distance to time difference between adjacent frames is calculated to obtain the instantaneous velocity sequence.

[0046] The variance of the instantaneous velocity sequence is calculated and its square root is taken to obtain the velocity fluctuation value.

[0047] Secondly, the present invention provides an artificial intelligence-based smart park target recognition system, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned artificial intelligence-based smart park target recognition method is implemented.

[0048] The beneficial effects are as follows: The core innovation of this invention lies in proposing a cross-blind zone target recognition mechanism based on dynamic weight fusion. Its breakthrough lies in constructing a feature effectiveness decay model, which non-linearly reduces the weight of visual features according to dwell time and environmental complexity; simultaneously, it introduces a probabilistic prediction model including a time diffusion term, automatically widening the spatiotemporal constraint range as dwell time increases. This strategy of dynamically complementing visual features and spatiotemporal logic effectively solves the problems of appearance failure and large motion prediction deviations caused by long dwell times, significantly improving recognition accuracy. Attached Figure Description

[0049] Figure 1 This is a flowchart of a smart park target recognition method based on artificial intelligence.

[0050] Figure 2 This is a dynamic decay curve of the feature confidence coefficient as a function of the complexity of the blind zone environment.

[0051] Figure 3 This diagram illustrates the dynamic switching mechanism between visual features and spatiotemporal probability weights. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] An embodiment of the AI-based smart park target recognition method provided by this invention:

[0054] like Figure 1 As shown, the AI-based smart park target recognition method includes the following steps:

[0055] S1 accesses the park's surveillance video stream, analyzes the camera topology to determine blind spot channels, and obtains the physical path length and light intensity index of blind spot channels through map mapping and historical image analysis. At the same time, it extracts the trajectory point sequence of the target within the visible area.

[0056] In one embodiment, a digital mapping of the physical world is first established. The system then accesses multiple surveillance video streams from the park and uses a target detection algorithm to extract targets from each frame. For each detected target, its trajectory point sequence within the camera's field of view is recorded. ,in As coordinates, For timestamps.

[0057] Simultaneously, the camera topology of the park is predefined. The first and second cameras are assumed to be adjacent, and the non-visual area between them is defined as a blind spot. Two basic parameters of this blind spot are obtained:

[0058] Physical path length : Obtained through map surveying, for example, the tree-lined path connecting the office building and the canteen within the park, with a surveyed length of [length missing]. rice.

[0059] Light intensity variation index This is obtained by comparing the historical average brightness histograms of the first camera's exit and the second camera's entrance, and then normalizing the results to 0-1. For example, if the first camera is in a shadow area and the second camera is in a bright area, the calculated difference is normalized. .

[0060] By filtering the dataset and accurately calibrating the physical environment parameters, an accurate spatiotemporal topology model of the park can be established, providing reliable data support for subsequent algorithm analysis and thereby improving the system's ability to perceive basic environmental information.

[0061] S2 monitors the dwell time of the target after entering the blind zone. Based on the speed fluctuation of the target before entering the blind zone and the environmental complexity factors of the blind zone channel, the confidence weight of the visual features is calculated through a preset feature effectiveness attenuation model. The confidence weight decreases non-linearly with the increase of dwell time.

[0062] In one embodiment, when the target When the device disappears from the first camera and enters the blind spot, the system starts a timer to record the time it spends in the blind spot in real time. To assess the reliability of visual features after passing through blind spots, it is necessary to calculate the feature confidence decay coefficient. .

[0063] First, calculate the two process variables:

[0064] Speed ​​fluctuation value : Extract the target before entering the blind zone Calculate the standard deviation of the instantaneous velocity sequence of the frame. For example, if a pedestrian walks at varying speeds before entering a blind spot, calculate the standard deviation of their velocity. .

[0065] Environmental complexity factors : Combining the number of branch roads within the blind spot and the aforementioned changes in light intensity Perform the calculation; the formula is as follows: Assume a certain blind spot has a certain light index There is one branch road. Set the maximum base number of branch roads. ,but .

[0066] Next, the confidence weights are calculated using the feature effectiveness decay model. The formula is as follows:

[0067]

[0068] in, The basic stability constant is set to 10.0. This is the logarithmic safety constant, with a value of 3.0.

[0069] The calculation example is as follows:

[0070] Assuming the target remains in the blind spot Seconds, environmental complexity factors speed fluctuation value Substitute into the formula to calculate: Logarithmic term: The second term in the denominator: Sum of denominators: ;result: .

[0071] The results indicate that due to the long dwell time and complex environment, the confidence level of visual features drops to 0.11, and the system should significantly reduce its trust in visual features.

[0072] By incorporating environmental and behavioral variables to construct a dynamic decay model, the extent to which visual features become ineffective over time can be assessed, thereby avoiding blindly trusting visual matching results when features change drastically and obtaining more accurate feature weights.

[0073] S3, based on the instantaneous velocity of the target when entering the blind zone and the physical path length of the blind zone channel, uses a probabilistic prediction model that includes a time spread term to calculate the spatiotemporal arrival probability of the target reaching the next camera after the dwell time. The time spread term is used to evaluate the range of prediction error that increases over time.

[0074] In one embodiment, when visual features fail or their weight decreases, the probability of a target's appearance is inferred using physical motion laws. The system calculates the spatiotemporal arrival probability of the target reaching the next camera after a dwell time. .

[0075] The core formula is as follows:

[0076]

[0077] in, The instantaneous velocity of the target when it enters the blind spot. The physical length of the blind spot passage. The standard deviation of the path length is the baseline value. The time diffusion coefficient is 0.15.

[0078] The calculation example is as follows:

[0079] Continuing with the previous example, let's assume the target's entry velocity... meters per second, physical length of blind zone channel Meters, stay time Seconds, standard deviation baseline Meters, substitute into the formula to calculate: Calculate the numerator: ; Calculate the time diffusion term in the denominator: ; Calculate the total value in the denominator: Index section: ;result: .

[0080] This means that although the target arrives later than the theoretical time, the spatiotemporal arrival probability calculated by the system is 0.67 due to the introduction of the diffusion term. The system considers the match to be reasonable in spatiotemporal logic. Without the diffusion term, the denominator is only 5 and the exponent is -1, and the probability drops to 0.36, which can easily lead to missed detection.

[0081] By introducing a time diffusion mechanism, the uncertainty of the target position as it moves over time can be characterized, giving the prediction model physical fault tolerance and significantly improving the ability to capture targets with non-uniform motion.

[0082] S4. When the next camera captures a candidate target, the visual similarity between the candidate target and the original target is calculated. The visual similarity and the spatiotemporal arrival probability are dynamically weighted and fused using confidence weights to obtain the final matching score, thereby determining the target's identity.

[0083] In one embodiment, when a candidate target is captured by a next camera, its appearance features are extracted and its visual similarity is calculated with that of the original target. Final match score The calculation is as follows:

[0084]

[0085] in, This is the dimensional balance coefficient, with a value of 0.8, used to calibrate the numerical range difference between the probability value and the similarity value.

[0086] The calculation example is as follows:

[0087] Based on the data from the above steps: , Assuming visual similarity Calculate the first term: ; Calculate the second term: ;result: .

[0088] Without the dynamic weighting mechanism of this invention, if only average weighting is used, the score would be approximately... This invention outputs a better matching score through dynamic weight fusion, and is more inclined to identify the candidate target and the original target as the same subject. This result is consistent with the characteristics of actual scenarios: although the visual similarity of the target is low due to the influence of the blind spot environment, the motion logic reflected by the spatiotemporal arrival probability is fully reasonable.

[0089] Through dynamic weighted fusion and decision-making, the system can automatically switch to a judgment logic dominated by spatiotemporal features when visual features are unreliable, thereby achieving accurate association of targets across blind spots.

[0090] To visually verify the effectiveness of this solution, refer to Figure 2 and Figure 3 Please provide an explanation.

[0091] Figure 2 The graph shows the variation of feature confidence coefficient with blind zone dwell time, including curves representing low, medium, and high environmental complexity factors. It can be seen that the curves exhibit a smooth downward trend, and the higher the environmental complexity, the faster the curve decreases. This intuitively demonstrates that the algorithm can accelerate the elimination of unreliable visual features based on the severity of the environment, consistent with the design logic of step S2.

[0092] Figure 3 Showing The relationship between the two weights in the calculation formula is that, over time, the visual feature weight curve decreases while the spatiotemporal probability weight curve increases, and the two form a clear intersection point in the figure. This intersection point vividly reveals the core decision-making mechanism of the present invention: during the short-term residence phase, the system prioritizes trusting vision; after crossing the critical time point, the system smoothly transitions to prioritizing trusting spatiotemporal logic.

[0093] An embodiment of the AI-based smart park target recognition system provided by this invention:

[0094] The AI-based smart park target recognition system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned AI-based smart park target recognition method.

[0095] The AI-based smart park target recognition system also includes other components well-known to those skilled in the art, such as communication interfaces. Their setup and functions are known in the art and will not be described in detail here.

[0096] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained by such a computer-readable medium.

[0097] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for target identification of a smart park based on artificial intelligence, characterized in that, Includes the following steps: Access the park's surveillance video stream, analyze the camera topology to determine blind spot channels, and obtain the physical path length and light intensity index of blind spot channels through map mapping and historical image analysis. At the same time, extract the trajectory point sequence of the target within the visible area. The system monitors the dwell time of a target after it enters the blind zone. Based on the speed fluctuations of the target before entering the blind zone and the environmental complexity factors of the blind zone channel, the system calculates the confidence weight of visual features using a preset feature effectiveness attenuation model. The confidence weight decreases non-linearly with the increase of dwell time. Feature validity decay model: ; wherein is a confidence weight for the visual feature, is a base stability constant, is an environmental complexity factor, is a dwell time, is a speed fluctuation value before target entry, is a logarithmic safety constant; The calculation process of the environmental complexity factor includes: obtaining the number of branch roads in the blind zone and the maximum reference number of branch roads in the blind zone; calculating the ratio of the number of branch roads to the maximum reference number of branch roads; assigning preset weights to the light intensity change index and the ratio of the number of branch roads to the maximum reference number of branch roads respectively, and summing them to obtain the environmental complexity factor. Based on the instantaneous velocity of the target when entering the blind zone and the physical path length of the blind zone passage, a probabilistic prediction model including a time spread term is used to calculate the spatiotemporal arrival probability of the target reaching the next camera after the dwell time. In the formula, For the probability of arrival in spacetime, The instantaneous speed of the target when it enters the blind spot. The physical length of the blind spot passage. The standard deviation of the path length is the baseline value. The time diffusion coefficient is used to assess the range of prediction error as time increases. When the next camera captures a candidate target, the visual similarity between the candidate target and the original target is calculated. The visual similarity and the spatiotemporal arrival probability are dynamically weighted and fused using confidence weights to obtain the final matching score, thereby determining the target's identity.

2. The AI-based smart park target recognition method according to claim 1, characterized in that, Obtain the physical path length and illumination change intensity index of the blind zone channel, including: The non-visual area between two adjacent cameras in the camera topology is defined as the blind zone channel. The physical path length of the blind spot passage is obtained through map surveying; By comparing the historical average brightness histograms of the exit of the upper camera and the entrance of the lower camera in two adjacent cameras, and normalizing the comparison results, the intensity index of light change is obtained.

3. The AI-based smart park target recognition method according to claim 1, characterized in that, Calculating the confidence weights of visual features includes: Extract the instantaneous velocity sequence of the target within a preset number of frames before entering the blind zone, calculate the standard deviation of the instantaneous velocity sequence, and obtain the velocity fluctuation value; The environmental complexity factor is calculated by combining the number of branch roads in the blind zone and the intensity index of light change. By substituting the speed fluctuation value, environmental complexity factor, and residence time into the feature effectiveness decay model, the confidence weight is calculated.

4. The smart park target recognition method based on artificial intelligence according to claim 1, characterized in that, Calculate the spatiotemporal arrival probability of the target reaching the next camera after the dwell time, including: Obtain the instantaneous velocity of the target when it enters the blind zone; Calculate the absolute deviation between the theoretical travel distance and the physical path length of the target during the dwell time; The standard deviation of the path length is corrected using the time diffusion term to obtain the dynamic error range; The spatiotemporal arrival probability is calculated based on the ratio of absolute deviation to dynamic error range.

5. The smart park target recognition method based on artificial intelligence according to claim 1, characterized in that, The formula for calculating the final matching score is: In the formula, For the final match score, The confidence weights for visual features are... The visual similarity between the candidate target and the original target. For the probability of arrival in spacetime, This is the dimensional balance coefficient.

6. The smart park target recognition method based on artificial intelligence according to claim 3, characterized in that, The calculation process for velocity fluctuation values ​​includes: Acquire the last preset number of video frames before the target enters the blind zone; The coordinates of the center point of the target in each frame are extracted using a target detection algorithm. The ratio of displacement distance to time difference between adjacent frames is calculated to obtain the instantaneous velocity sequence. The variance of the instantaneous velocity sequence is calculated and its square root is taken to obtain the velocity fluctuation value.

7. A smart park target recognition system based on artificial intelligence, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent park target recognition method based on artificial intelligence as described in any one of claims 1-6 is implemented.

Images