Artificial intelligence-based pedestrian trajectory tracking method

Abstract

The present application relates to the field of artificial intelligence tracking technology, and discloses a pedestrian trajectory tracking method based on artificial intelligence.The method comprises collecting multi-source sensor pedestrian movement feature data stream, extracting initial trajectory key points as tracking reference points;dividing associated historical position data into a high-confidence trajectory segment set and a to-be-verified trajectory segment set, and discarding the to-be-verified set;arranging the high-confidence set in descending order according to the number of trajectory points, selecting the first several segments as optimal trajectory segments, and calculating the proportion;comparing the proportion with a dynamic screening threshold, outputting the high-confidence set as an initial tracking result, or outputting a trajectory checking instruction generated by comparing trajectory data;in response to the checking instruction, dividing the space grid of the optimal trajectory segment, and outputting an optimization instruction in combination with historical trajectory features;updating the reference points based on the optimization instruction, and screening and outputting secondary divided trajectory data;finally, obtaining parameters to generate a tracking efficiency evaluation coefficient, and starting a parameter correction strategy when the coefficient is lower than an optimization threshold.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence tracking technology, specifically to a method for tracking pedestrian trajectories based on artificial intelligence. Background Technology

[0002] In scenarios such as urban traffic management, public safety and security, and smart shopping mall customer flow analysis, pedestrian trajectory tracking is a crucial technical means to achieve intelligent operation. With the widespread adoption of surveillance equipment and the development of sensor technology, multi-source sensors have become the primary tools for collecting pedestrian movement data. These sensors include video cameras, infrared detectors, and millimeter-wave radar, capable of acquiring multi-dimensional feature data streams such as pedestrian position, speed, and direction of movement. However, current pedestrian trajectory tracking methods face numerous challenges in practical applications.

[0003] Data acquired from multiple sensors often contains noise interference. For example, video cameras are susceptible to changes in lighting and obstructions, while infrared detectors may experience signal drift in complex environments, leading to insufficient accuracy in the extracted initial trajectory key points. If tracking is directly based on these error-laden key points, problems such as trajectory deviation and breakage can easily occur, affecting the reliability of subsequent trajectory analysis.

[0004] Existing methods lack an effective trajectory segment classification mechanism when processing historical location data. Most technologies include all historical trajectory segments in the tracking calculation without distinguishing between the confidence levels of the segments. Some low-confidence trajectory segments may have been generated due to accidental interference during data collection or sudden behaviors such as pedestrian pauses or turns. Retaining these segments and including them in the trajectory calculation would increase the amount of data processing, reduce the overall accuracy of trajectory tracking, and even lead to trajectory misjudgments, such as confusing the trajectories of different pedestrians.

[0005] Current trajectory selection and optimization strategies lack dynamic adjustment capabilities. When selecting key trajectory segments, fixed thresholds or single selection criteria are often used, failing to flexibly adjust based on actual data volume and trajectory distribution characteristics. For example, when pedestrian density is high in a certain area and the number of trajectory segments increases significantly, a fixed selection threshold may lead to the rejection of many valid trajectory segments or the retention of too many redundant segments, affecting tracking efficiency. Furthermore, in the trajectory optimization stage, existing methods often rely on simple spatial interpolation or smoothing, failing to fully incorporate the distribution patterns of historical trajectories. This makes it difficult to accurately correct abnormal trajectory segments, resulting in deviations between the optimized trajectory and the actual pedestrian movement path.

[0006] Current technologies lack effective evaluation and feedback mechanisms for tracking efficiency. Most methods focus only on the final accuracy of trajectory tracking, without establishing a scientific evaluation index system to measure the efficiency of the entire tracking process, such as trajectory segment processing speed and the impact of parameter adjustments on tracking results. When tracking efficiency declines, the root cause cannot be identified and parameter correction strategies initiated in a timely manner, leading to the system operating in an inefficient state for extended periods. This is especially problematic in large-scale pedestrian scenarios, where data processing delays are likely to occur, failing to meet the demands of real-time tracking. These issues collectively limit the application effectiveness of pedestrian trajectory tracking technology in real-world scenarios, necessitating a more efficient and accurate trajectory tracking method to address these shortcomings. Summary of the Invention

[0007] The purpose of this invention is to provide an artificial intelligence-based pedestrian trajectory tracking method to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a pedestrian trajectory tracking method based on artificial intelligence, the method comprising:

[0009] Collect pedestrian movement feature data streams from multiple source sensors and extract key points of the initial trajectory as tracking reference points;

[0010] The historical location data associated with the key points of the initial trajectory are divided into a set of high-confidence trajectory segments and a set of trajectory segments to be verified, and the set of trajectory segments to be verified is discarded.

[0011] Receive a set of high-confidence trajectory segments, sort them in descending order by the number of trajectory points, select the top few segments as preferred trajectory segments, and calculate the proportion of preferred trajectory segments in the set of high-confidence trajectory segments;

[0012] When the proportion is lower than the dynamic filtering threshold, a set of high-confidence trajectory segments is output as the initial tracking result; otherwise, the comparison trajectory data is output and a trajectory verification instruction is generated.

[0013] In response to the trajectory verification command, the preferred trajectory segment is divided into spatial grids, and the trajectory optimization command is output based on the historical trajectory distribution characteristics.

[0014] The tracking reference point is updated based on the trajectory optimization command, and the trajectory data of the secondary division is filtered and output.

[0015] Obtain the proportion of preferred trajectory segments, the rate of change in the number of trajectory points, and the number of trajectory segments before and after instruction execution to generate a tracking efficiency evaluation coefficient;

[0016] When the tracking efficiency evaluation coefficient is lower than the optimization threshold, the tracking parameter correction strategy is activated.

[0017] Preferably, a spatiotemporal filtering algorithm is used to preprocess the sensor data stream before extracting the key points of the initial trajectory;

[0018] The pedestrian movement feature data stream is spatiotemporally encoded using a graph neural network, and the encoding results are fused with kinematic model parameters to generate a new trajectory prediction output.

[0019] Preferably, the specific steps for segmenting historical location data include:

[0020] Construct a trajectory confidence region centered on the key points of the trajectory;

[0021] When the number of trajectory points within the confidence region is less than the preset trajectory density threshold, it is included in the set of trajectory segments to be verified.

[0022] Conversely, they are classified into the set of high-confidence trajectory segments;

[0023] When executing trajectory optimization instructions, if the growth rate of the number of trajectory points in the set of trajectory segments to be verified exceeds a preset percentage, it will be transferred to a set of high-confidence trajectory segments.

[0024] Preferably, the adjustment steps for the dynamic screening threshold include:

[0025] Obtain the ratio of the number of trajectory segments to be verified to the number of high-confidence trajectory segments in the current scenario;

[0026] When the ratio exceeds the upper limit threshold of trajectory density or falls below the lower limit threshold of trajectory density;

[0027] Calculate the average trajectory point density for all trajectory segments;

[0028] When the average trajectory point density is greater than the product of the upper limit threshold for trajectory density and the ratio, the upper limit threshold for trajectory density is increased.

[0029] When the average trajectory point density is less than the product of the trajectory density lower limit threshold and the ratio, the trajectory density lower limit threshold is lowered.

[0030] Preferably, the steps for selecting the preferred trajectory segment include:

[0031] Retrieve the average effective trajectory length of historical tracking tasks;

[0032] When the total number of high-confidence trajectory segments does not exceed the average effective trajectory length, the high-confidence trajectory segment set is output directly.

[0033] Conversely, if the number of trajectory points is low, the trajectory segments with an equal length to the average effective trajectory length are selected as the preferred trajectory segments in descending order.

[0034] If the continuous tracking time does not reach the preset period, retrieve the set of trajectory segments to be verified as supplementary trajectory data output.

[0035] Preferably, the steps for generating trajectory verification instructions include:

[0036] The conversion rate of unverified trajectory segments into high-confidence trajectory segments in historical tracking tasks was statistically analyzed.

[0037] The number of trajectory points that are merged from the current set of high-confidence trajectory segments and the set of trajectory segments to be verified;

[0038] When the calculated trajectory anomaly risk index exceeds the preset risk threshold, the set of high-confidence trajectory segments is directly output.

[0039] Conversely, a trajectory verification command is generated.

[0040] Preferably, the evaluation steps for spatial grid generation include:

[0041] The preferred trajectory segment is divided into multi-scale grids, and the difference between the maximum and minimum number of trajectory points in each grid cell is calculated.

[0042] When the difference exceeds the spatial uniformity threshold, the grid cells are re-divided;

[0043] Record the division scale parameters for qualified grid cells;

[0044] A grid division score is generated by combining the distribution density and positional deviation of trajectory points in each grid cell;

[0045] The scale parameter corresponding to the highest grid division score is selected as the trajectory optimization instruction.

[0046] Preferably, the calculation steps for the tracking efficiency evaluation coefficient include:

[0047] The rate of change of the number of preferred trajectory segments and trajectory points before and after the execution of the fusion instruction;

[0048] Overlaying the trajectory anomaly risk index with preset weight parameters;

[0049] When the total number of high-confidence trajectory segments does not exceed the average effective trajectory length, the normalization coefficient is used instead of the number of trajectory points for calculation.

[0050] Preferably, the execution steps of the tracking parameter correction strategy include:

[0051] Grid scale parameters are selected in descending order of grid division scores, and the tracking efficiency evaluation coefficients for each scale parameter are calculated.

[0052] Select the grid scale parameter that causes the tracking efficiency evaluation coefficient to fall below the optimization threshold for the first time;

[0053] Configure this grid scale parameter as the final tracking parameter.

[0054] Preferably, the steps for updating the tracking benchmark include:

[0055] Map the secondary segmented trajectory data to spatial grid cells;

[0056] The coordinates of key points in the initial trajectory are corrected based on the density distribution of trajectory points within the grid cell.

[0057] The updated tracking reference points are generated by integrating the correction results of continuous grid cells using a time sliding window.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] In the data preprocessing stage, this method first processes the pedestrian movement feature data stream acquired by multi-source sensors, extracting initial trajectory key points as tracking reference points. Simultaneously, it divides the historical location data associated with these initial trajectory key points into a high-confidence trajectory segment set and a set of trajectory segments to be verified, discarding the latter. This process accurately filters out highly reliable trajectory data, removing low-confidence trajectory segments caused by sensor noise, environmental interference, or sudden pedestrian behavior. This reduces the interference of invalid data on subsequent tracking calculations, ensuring the accuracy of trajectory tracking from the data source and avoiding problems such as trajectory deviation, breakage, or confusion caused by low-quality data. This allows subsequent trajectory analysis to be based on more reliable data.

[0060] In the trajectory segment selection and processing stage, after receiving the set of high-confidence trajectory segments, this method sorts them in descending order by the number of trajectory points and selects the top few segments as preferred trajectory segments. Simultaneously, it calculates the proportion of preferred trajectory segments in the set of high-confidence trajectory segments and, based on the comparison between this proportion and a dynamic selection threshold, flexibly outputs different tracking results or generates trajectory verification instructions. This dynamic selection mechanism can adjust the processing strategy according to the actual distribution of high-confidence trajectory segments. When the proportion of preferred trajectory segments is low, it indicates that the overall distribution of high-confidence trajectory segments is relatively scattered. In this case, outputting the set of high-confidence trajectory segments as the initial tracking result avoids over-selection leading to the loss of effective data. When the proportion is high, generating trajectory verification instructions further optimizes the process, ensuring that the selected trajectory segments are both representative and cover key tracking information, balancing the accuracy of trajectory tracking with the rationality of data processing, and avoiding the limitations of traditional fixed-threshold selection methods.

[0061] In the trajectory optimization and update stage, after responding to the trajectory verification command, the preferred trajectory segment is spatially gridded, and a trajectory optimization command is output based on historical trajectory distribution characteristics. The tracking reference point is then updated based on this command, and the secondary-divided trajectory data is filtered and output. Spatial grid division refines the preferred trajectory segment to a more precise spatial range, while combining it with historical trajectory distribution characteristics fully utilizes past pedestrian movement patterns to identify and correct abnormal parts in the current trajectory segment. For example, trajectory breaks caused by temporary occlusion are appropriately filled in, and trajectory segments deviating from the normal movement path are adjusted. This makes the updated tracking reference point more closely match the actual movement state of the pedestrian, and the secondary-divided trajectory data more accurately reflects the pedestrian's movement trajectory, further improving the accuracy of trajectory tracking and making the trajectory results more in line with the needs of practical application scenarios.

[0062] To ensure tracking efficiency, this method obtains the proportion of preferred trajectory segments, the rate of change in the number of trajectory points, and the number of trajectory segments before and after command execution to generate a tracking efficiency evaluation coefficient. When this coefficient falls below the optimization threshold, a tracking parameter correction strategy is initiated. This evaluation and correction mechanism can monitor the efficiency of the entire trajectory tracking process in real time, comprehensively judge the operating status of the tracking system through multi-dimensional indicators, and promptly initiate parameter corrections once a decline in efficiency is detected. For example, adjusting the criteria for the number of trajectory segments selected and optimizing the accuracy of spatial grid division ensures that the system always maintains a high-efficiency operating state, avoids processing delays when pedestrian density changes or sensor data fluctuates, and meets the real-time requirements of trajectory tracking in different scenarios. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the working principle of the pedestrian trajectory tracking method based on artificial intelligence described in this invention.

[0064] Figure 2 A flowchart for trajectory segment division and optimization;

[0065] Figure 3 This is a flowchart for dynamically adjusting the filtering threshold;

[0066] Figure 4 A flowchart generated for the trajectory verification command;

[0067] Figure 5 A flowchart for tracking the execution of parameter correction strategies. Detailed Implementation

[0068] 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 embodiments of the present invention, and not all embodiments. 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.

[0069] Please see Figure 1 This invention provides a pedestrian trajectory tracking method based on artificial intelligence, the method comprising:

[0070] The system collects real-time pedestrian movement feature data streams from multiple sensors, including but not limited to position, speed, and direction information acquired by devices such as visual sensors, millimeter-wave radar, and lidar. After preprocessing the raw data stream using a spatiotemporal filtering algorithm, a sliding time window is used to extract trajectory key points with significant motion characteristics as tracking reference points. Historical location data associated with each key point is divided into a high-confidence trajectory segment set and a set of trajectory segments to be verified using a density clustering algorithm, with the latter being discarded. Upon receiving the high-confidence trajectory segment set, the system first sorts the segments in descending order based on the number of trajectory points they contain, selecting the top N segments as preferred trajectory segments. The value of N is dynamically adjusted according to the current scene. The proportion of preferred trajectory segments in the high-confidence trajectory segment set is calculated. If this proportion is lower than the dynamic filtering threshold, the high-confidence trajectory segment set is directly output as the initial tracking result; otherwise, the trajectory comparison module is activated, outputting the comparison trajectory data and generating a trajectory verification command. Upon responding to the command, the trajectory verification module performs multi-scale spatial grid division on the preferred trajectory segments and generates trajectory optimization commands based on the distribution characteristics in the historical trajectory database. The system updates the tracking benchmark based on optimization instructions and filters and outputs the secondary segmented trajectory data. The tracking evaluation module calculates in real time the proportion of preferred trajectory segments, the rate of change in the number of trajectory points, and the change in the number of trajectory segments before and after instruction execution, generating a tracking efficiency evaluation coefficient. When this coefficient is lower than a preset optimization threshold, a tracking parameter correction strategy is automatically triggered.

[0071] Example 1: See Figure 2 Taking a city intersection surveillance scenario as an example, the deployed visual sensors capture pedestrian positions at 30 frames per second, the millimeter-wave radar updates velocity vectors at a frequency of 10Hz, and the lidar provides precise 3D coordinates 20 times per second. The raw data stream first enters the preprocessing channel: a Kalman filter removes noise from the radar velocity data, a temporal convolutional network smooths the visual position information, and a bidirectional LSTM network compensates for data loss from the lidar in rainy or foggy weather. After three levels of processing, the time alignment error of the data stream is controlled within 50 milliseconds.

[0072] The preprocessed data is input into a graph neural network for spatiotemporal encoding. This network contains 12 graph convolutional layers, with the output of each layer passed to the next layer via residual connections. The third layer extracts the relative positional relationships between pedestrians, the sixth layer captures group motion patterns, and the ninth layer identifies individuals with abnormal movements. The 128-dimensional feature vector output by the network is concatenated with the state vector of the kinematic model in a fusion layer. The kinematic model employs an improved constant velocity-constant steering rate algorithm, and its state vector contains six parameters: position, velocity, acceleration, and steering angle. These parameters are updated every 200 milliseconds using an adaptive particle filter. A weighted concatenation mechanism is used for feature fusion, with graph neural network features accounting for 70% of the weight and kinematic parameters accounting for 30%, ultimately generating a trajectory prediction output containing 134-dimensional features.

[0073] During the historical location data segmentation phase, when a pedestrian is detected turning right, an elliptical confidence region is constructed centered on the current trajectory key point. The major axis of the ellipse maintains a 15-degree deviation angle from the pedestrian's movement direction, the minor axis is fixed at 1.2 meters, and the major axis is dynamically calculated based on the real-time speed v using the formula 1.2 + 0.5v (meters). The trajectory point density within the confidence region is estimated using a Gaussian kernel function with a kernel bandwidth of 0.8 meters. When the system detects that the density value of a certain area is lower than 0.3 points / square meter for five consecutive sampling periods, the trajectory segment in that area is marked as pending verification.

[0074] The determination of a high-confidence trajectory segment requires the simultaneous fulfillment of three conditions: a density value consistently higher than 0.5 points / square meter, an interval between adjacent trajectory points of less than 200 milliseconds, and a change rate of motion direction not exceeding 15 degrees within a 1-second window. When a pedestrian completes a turning action, the system detects that the number of trajectory points in the set of trajectory segments to be verified has increased by 45% within 3 seconds, exceeding the preset threshold of 20%, and then triggers an automatic conversion mechanism: the data of 32 trajectory points collected during that period are transferred to the high-confidence trajectory segment set, while the confidence scores of the key points of the trajectory are updated.

[0075] During the execution of trajectory optimization instructions, when a trajectory segment to be verified at a shopping mall entrance is found to have increased by 210% within 10 seconds, an emergency conversion procedure is automatically initiated: First, the temporal continuity of these trajectory points is verified; after confirming there are no time jumps, the spatial distribution consistency is checked; finally, a set conversion operation is performed, and the average density value of the high-confidence trajectory segment is recalculated. The timestamps and spatial coordinate information of the original data are preserved during the conversion process to ensure that the integrity of the trajectory is not affected.

[0076] For crowded scenarios in subway stations, the system employs a layered processing strategy: maintaining standard confidence region parameters in the main passage area, while reducing the minor axis of the ellipse to 0.8 meters in narrow areas like ticket gates to improve discrimination accuracy. When a passenger is detected lingering in front of the turnstile, the system automatically extends the time window to 500 milliseconds and reduces the rate of change of direction threshold to 8 degrees / second to avoid misjudging normal queuing behavior as abnormal trajectories. During the morning rush hour, the system recorded 127 triggers of set conversion operations per hour, successfully integrating the fragmented movement trajectories of waiting passengers into continuous high-confidence trajectory segments.

[0077] Example 2: See Figure 3 A dynamic threshold adjustment mechanism operates in large transportation hub scenarios. Taking a train station exit as an example, the system scans the current trajectory data status approximately every 5 seconds. When a train arrives, the pedestrian density in the exit passage increases sharply. The system detects that the ratio of the number of trajectory segments to be verified to the number of high-confidence trajectory segments reaches 1.62, exceeding the preset upper limit threshold of 1.5. At this point, the threshold adaptive module is activated, calculating the average trajectory point density of all current trajectory segments to be 28 points / square meter. Since this value is greater than the product of the upper limit threshold and the ratio (1.5 × 1.62 = 2.43), the system increases the upper limit threshold for trajectory density to 1.6 in 0.1 increments. The adjusted threshold remains locked for the next 10 seconds, during which time the system continues to monitor changes in the ratio.

[0078] When passenger flow stabilizes, the ratio drops to 0.25 and remains so for two detection cycles. This value is below the lower threshold of 0.3, triggering a threshold adjustment procedure. The system recalculates the average trajectory point density to be 19 points / square meter. This value is less than the product of the lower threshold and the ratio (0.3 × 0.25 = 0.075), so the lower threshold is lowered by 0.05 to 0.25. The adjusted threshold range is limited to the effective range to avoid parameter failure due to extreme conditions.

[0079] The preferred trajectory segment selection module operates in the transfer passage scenario. The system retrieves the average effective trajectory length of the most recent 100 tracking tasks from the historical database, and this value is weighted to 15 segments. The current high-confidence trajectory segment set contains 42 trajectories, exceeding the benchmark value and triggering a hierarchical filtering mechanism. The first level of filtering retains trajectory segments with more than 30 trajectory points (twice the average), resulting in 8 high-quality trajectories. The second level of filtering supplements 7 trajectories from the remaining 34 trajectories based on spatial distribution uniformity: prioritizing 3 segments from east-west passages, 2 segments from north-south connecting corridors, and 2 segments from stairwell areas, ensuring that the final output of preferred trajectory segments is exactly equal to 15 segments.

[0080] When the tracking time is less than the preset period (e.g., the first 20 seconds after system startup), a supplementary data mechanism is activated. Taking a subway turnstile as an example, newly emerging passenger trajectories have not yet reached the standard tracking period. The system scans the set of trajectory segments to be verified and selects 5 trajectories that meet the minimum density requirement (0.4 trajectory points per square meter). These trajectory segments, after spatial location verification, are added to the output queue as supplementary data, with their total amount controlled at 18% of the total output (below the 20% upper limit). Supplementary trajectories are marked with special identifiers and use differentiated confidence parameters in subsequent processing.

[0081] When applied to an open-air plaza scenario, the dynamic threshold adjustment demonstrated adaptability. During the spring event, significant fluctuations in pedestrian flow occurred, triggering seven threshold adjustments within one hour: three increases in the upper limit due to crowd gathering (cumulative increase of 0.3), and four decreases in the lower limit due to crowd dispersal (cumulative decrease of 0.15). Each adjustment was followed by a 10-second stabilization period, during which parameter changes were frozen while data fluctuations were continuously recorded. After the stabilization period, the system determined whether a secondary adjustment was necessary based on newly collected data.

[0082] When selecting the optimal trajectory segments, an uneven distribution of trajectory points was encountered in the central area of ​​the square. The high-confidence trajectory segment set contained 38 segments, but 70% were concentrated around the fountain. The system performed a spatially balanced filtering: first, the top five trajectory segments with the most trajectory points in the fountain area were selected, and then the top ten segments with the most trajectory points in descending order were selected from the edge areas. During the selection process, the spatial coverage uniformity was calculated in real time. When a missing trajectory segment was detected in the northwest area, two trajectory segments from that area were added from the suboptimal queue. The final output of 15 trajectory segments covered the six main areas of the square, with the error in the proportion of each area controlled within 5%.

[0083] In scenarios where initial data is insufficient, the supplementation mechanism plays a crucial role. On the first day of operation of a newly built bus hub, the system lacked historical data for reference. During the first 10 minutes of operation, when the set of high-confidence trajectory segments was less than 15 segments, the system automatically increased the supplementation ratio of trajectory segments to be verified to 22% (a temporary relaxation of 3 percentage points). Supplemented data must pass dual verification: checking the continuity of trajectory points in the time dimension (missing rate less than 15%) and verifying the consistency of movement direction in the spatial dimension (angle deviation less than 25 degrees). These temporarily supplemented trajectory segments are marked with a yellow warning label and will be prioritized for replacement when sufficient data becomes available later.

[0084] Example 3: See Figure 4The trajectory verification instruction generation module operates in a commercial center scenario. The system calculates the conversion rate of trajectory segments to be verified into high-confidence trajectory segments within the last 10 minutes, updating this data every 30 seconds via a sliding time window. On a certain weekday afternoon, the system detected that the conversion rate in the escalator area remained below 0.15. Simultaneously, the current high-confidence trajectory segment set contained 1200 trajectory points, while the set to be verified contained 900 trajectory points, resulting in a ratio of 1.33. The environmental interference factor, calculated based on WiFi signal strength and millimeter-wave reflection intensity, was measured at 0.68 in areas with dense metal structures.

[0085] The risk prediction model uses a three-layer fully connected network to process these inputs: the input layer receives three standardized feature vectors, the hidden layer contains eight neurons using the ReLU activation function, and the output layer generates a trajectory anomaly risk index using the sigmoid function. This index is calculated using the following formula:

[0086] ;

[0087] in: This indicates an abnormal trajectory risk index. Represents the sigmoid function. This is a standardized value for the historical conversion ratio. This represents the ratio of the current number of trajectory points (number of high-confidence caissons / number of caissons to be verified). It is an environmental disturbance factor. , , These are the network weight parameters. This is the bias term. When the escalator area measures R=0.82 during peak hours, exceeding the preset risk threshold of 0.7, the system directly outputs a set of high-confidence trajectory segments.

[0088] When the risk index is below the threshold, such as when R=0.45 is measured in the open area of ​​the atrium, the verification instruction generator is activated. The module output includes three types of control parameters: the grid division accuracy requirement is set to 0.8m (automatically adjusted according to spatial complexity), the historical data backtracking depth is set to the trajectory of the most recent 5 minutes, and the optimization tolerance threshold is set to 15%. These parameters are synchronously transmitted to the spatial processing unit via the configuration bus.

[0089] When implementing spatial gridding in the dining area corridor, the system performed multi-scale analysis on the preferred trajectory segments. The initial outer rectangle size was 12m × 6m. The system attempted three grid scales: 0.5m × 0.5m generating 288 cells, 1m × 1m generating 72 cells, and 2m × 2m generating 18 cells. At the 0.5m scale, grid cell number 45 was found to contain 8 trajectory points, while the adjacent cell number 46 contained only 1 point. The range of 7 exceeded the preset threshold of 3 for this scale, and this scheme was rejected. In the 1m scale scheme, the largest grid cell contained 15 points (near the food pick-up area), and the smallest cell contained 3 points (near the fire escape). The range of 12 exceeded the threshold of 5, and this scheme was also rejected.

[0090] The 2m×2m scale scheme passed the initial inspection: the largest cell contained 28 points, the smallest cell contained 17 points, and the range of 11 was below the threshold of 8. The system calculated the spatial uniformity score of the scheme: the measured density variance of each cell was 4.7 (6.5 out of 10), and the trajectory continuity score: 85% of the trajectory points were detected to be located in continuous grids (8 points). The overall score = 6.5 × 0.6 + 8 × 0.4 = 7.1. This score was written into the trajectory optimization command, and the standard distribution density range of each grid cell was marked as 17-28 points / 4m², with an allowable deviation of ±3 points.

[0091] In the scene depicting the movie theater exit, the system detected a unique distribution pattern: six main passageways arranged radially. The grid generation module automatically switched to a fan-shaped grid mode, dividing the space into 30-degree angle sectors. During the sector selection evaluation, the range calculation was adjusted to compare sectors within the same ring: the maximum density difference in the inner ring (0-10 meters) sector was 5 points, and the difference in the middle ring (10-20 meters) was 8 points, both below the corresponding thresholds. The system recorded the division scale parameters for each sector (ring width 10 meters), and the final grid generation score reached 8.3, making it the optimal solution for this scene.

[0092] When applied to escalator areas, the system activates a tilted grid generation mode, using a 35-degree escalator tilt angle as a reference to establish a rotating coordinate system and generate parallelogram-shaped grid cells. At a scale of 1.2m × 1.2m, the system detected a maximum of 22 trajectory points within each grid cell (running section) and a minimum of 18 (transition section), a difference of 4, which is below the threshold of 5. In spatial uniformity calculation, after coordinate transformation of the tilted grid, the system measured a density variance of 3.1, a trajectory continuity score of 9.2, and a comprehensive score of 8.6, which was then written into the optimization command.

[0093] During the execution of trajectory optimization instructions, when the system detects a sudden drop in light intensity of 300 lux in a certain area, it automatically adds supplementary instructions: temporarily expanding the allowable deviation of the standard distribution density of the grid cells to ±5 points, and extending the historical data backtracking depth from 5 minutes to 8 minutes. These dynamically adjusted parameters are transmitted through the instruction extension field to ensure that the spatial processing unit adapts to environmental changes in real time.

[0094] Example 4: See Figure 5 The tracking efficiency evaluation system operates in an airport terminal building scenario. Forty-seven multi-source sensor nodes are installed in the departure hall on the third floor of the terminal, collecting passenger movement trajectory data in real time. The system generates a tracking efficiency evaluation coefficient every 15 seconds. This coefficient is calculated based on real-time monitoring data in three dimensions: changes in the number of preferred trajectory segments before and after command execution, fluctuations in the number of trajectory points, and a trajectory anomaly risk index. Table 1 shows the evaluation parameter records for five consecutive periods from 10:00 AM to 10:15 AM on a certain morning.

[0095] Table 1: Record of tracking efficiency evaluation parameters in the departure hall of the terminal.

[0096] ;

[0097] In the check-in area, the system detected that the evaluation coefficient was below the optimization threshold of 0.65 for three consecutive cycles. The coefficient of 0.73 in the first cycle was within the normal range, the coefficient of 0.68 in the second cycle triggered an early warning, and the coefficient of 0.61 in the third cycle officially activated the parameter correction process. The system retrieved the scale parameters of the 20 most recent valid grid divisions from the grid division record library and formed a candidate queue by arranging them in descending order of score. The top five parameters in the queue were 1.8m, 1.5m, 1.2m, 0.9m, and 0.6m, with corresponding historical scores of 8.7, 8.5, 8.3, 8.1, and 7.9, respectively.

[0098] The system first applied parameters at a 1.8m scale, recalculating the evaluation coefficient to 0.67, still below the threshold but close to the critical value. Next, the system tested at a 1.5m scale, where the coefficient rose to 0.69, meeting the requirements, but the system continued to verify smaller scales. When testing at a 1.2m scale, the coefficient broke the threshold for the first time, reaching 0.72, and this parameter was determined to be the current optimal solution. The system set 80% of the 1.2m parameter (0.96m) as the new baseline for the grid generation module, while adjusting the dynamic filtering threshold to 1.2 times the original value, and extending the trajectory verification command generation cycle from 30 seconds to 39 seconds.

[0099] When flights arrive in concentrated periods, passenger trajectories exhibit a high-density clustering characteristic. The system recorded 42 trajectory segments before execution and 45 after execution, with a trajectory point change of +89, a risk index of 0.31, and a calculated evaluation coefficient of 0.71. Although the coefficient met the target, the system detected a sudden increase in the amount of trajectory point change and thus initiated a stability check procedure. The coefficient fluctuations for three consecutive periods were +0.03, -0.01, and +0.02, respectively, with fluctuation amplitudes all less than 5%, confirming that the parameter settings were stable and effective.

[0100] When abnormal fluctuations in the assessment coefficient occurred at the international arrivals channel, a sharp drop in the coefficient to 0.58 was detected within the 10:30:00 period. Investigation revealed that customs checks were causing prolonged passenger wait times. The system automatically switched to a low-speed mode: the time sliding window was extended from the standard 3 seconds to 5 seconds, the grid scale parameter was temporarily relaxed to 1.5 times the baseline value (1.44m), and the risk index weight of the assessment coefficient was reduced from 30% to 20%. After the adjustment, the coefficient gradually recovered to 0.64, 0.66, and 0.69 within three periods, and the system maintained the new parameters for continued monitoring.

[0101] In the passageway connecting the corridor bridges, the system handles long-distance tracking scenarios. The number of preferred trajectory segments increased from 18 before execution to 24 after execution, resulting in a trajectory point change of +72, a risk index of 0.25, and a calculated evaluation coefficient of 0.74. Since the total number of high-confidence trajectory segments in this area has consistently been lower than the average effective trajectory length (15 segments), the system uses a normalized coefficient to replace the calculation. The ratio of the current number of trajectory points to the historical average is taken as 1.15, and combined with other parameters, a final evaluation coefficient of 0.77 is generated. The system determines that the parameter configuration in this area is reasonable, and only fine-tunes the default grid division scale from 0.96m to 1.04m to adapt to changes in passage width.

[0102] When applied to security checkpoints, the metal detector gate caused attenuation of the millimeter-wave radar signal, and the system detected periodic fluctuations in the evaluation coefficients. A coefficient change sequence was recorded between 10:45:00 and 10:48:00: 0.70→0.62→0.65→0.63. The system activated its anti-interference mode: extending the evaluation calculation period from 15 seconds to 20 seconds and adding data smoothing processing; simultaneously limiting the parameter adjustment range, with a single scale parameter change not exceeding 0.2m. After three adjustment cycles, the coefficients stabilized in the 0.67-0.69 range, and the system continued to operate with the current configuration.

[0103] The terminal's commercial area exhibited consistently high evaluation coefficients (0.78-0.82), indicating that the current parameters might be overly conservative. The system performed reverse optimization: testing parameters at a 0.6m scale was selected from the end of the grid-based scoring queue, reducing the evaluation coefficient to 0.71, but still above the threshold. Further testing at a 0.75m scale resulted in a coefficient of 0.69, reaching the ideal equilibrium point. The system stored this parameter as a dedicated configuration for the commercial area in the regional parameter database and implemented geofencing for automatic switching. When passengers move from the commercial area to the departure lounge, the system automatically reverts to the default parameter configuration, ensuring real-time matching between the tracking strategy and environmental characteristics.

[0104] Example 5: The airport baggage carousel area demonstrates the complete process of tracking reference point updates. When passengers gather at carousel 3 to wait for their luggage, the system maps the secondary segmented trajectory data to the optimal spatial grid. This area uses a 1.5m × 1.5m grid scale, dividing a 120 square meter area into 64 cells. Each passenger's movement trajectory is decomposed and mapped to different grid cells. The system records three key features: calculating the average coordinates of all trajectory points within the cell as the density centroid, extracting the median of the timestamp sequence, and calculating the weighted average of the motion direction vectors.

[0105] When a passenger moves in front of the carousel, the system detects a 0.42-meter offset between the initial trajectory keypoint and the centroid of the corresponding grid density. This value exceeds 1 / 4 of the grid side length (0.375 meters), triggering a coordinate correction procedure. The correction algorithm first calculates the direction angle of the offset vector and finds a 15-degree deviation compared to the historical movement direction. The system then verifies this by combining the predicted trajectories from the previous three time windows: the first window (5 seconds ago) had a positional deviation of 0.2 meters, the second window (10 seconds ago) had a deviation of 0.18 meters, and the third window (15 seconds ago) had a deviation of 0.25 meters, all less than the tolerance threshold of 0.3 meters. After confirming the rationality of the correction, the new coordinates replace the original keypoints with an 85% weight.

[0106] When a passenger pushes a luggage cart at a speed of 0.8 m / s, the system activates a 5-second time window. The window slides in 0.5-second increments, integrating data from 12-15 consecutive grid cells in each window. Within window T1 (10:00:00-10:00:05), the system integrates 14 grid cells to generate a temporary reference point sequence. When the window slides to T2 (10:00:00.5-10:00:05.5), the newly added cell data coverage reaches 93%, and the system merges the old and new cell data to generate updated reference points.

[0107] A business traveler moves quickly through the waiting hall at a speed of 2.3 meters per second. The system automatically switches to a 1-second time window mode, integrating only 5-7 grid cells per window. Within a 10-second observation period, the system generates 10 consecutive reference points, with the time interval between each point precisely maintained at 1 second. The reference point data structure includes three-dimensional coordinates (longitude, latitude, and altitude), confidence scores, and timeliness indicators. Example of confidence score calculation: A point has a grid cell density of 0.7 points / square meter (contributing 42 points), a time series consistency score of 28 points (out of 30), a motion smoothness score of 8.5 points (out of 10), and a comprehensive score of 78.5 points.

[0108] When the system generates a reference point for a boarding gate that is about to close, it adds a timeliness identifier "EXP-08," indicating that the point is valid for the next 8 cycles (40 seconds). When the system detects a passenger accelerating towards the boarding gate, it dynamically shortens the timeliness to 4 cycles; if the passenger sits down in the rest area, the timeliness is extended to 12 cycles. The timeliness prediction model references three parameters: the rate of change of movement speed, the directional stability index, and the environmental congestion coefficient.

[0109] If a passenger is detected lingering in the coffee shop for more than 3 minutes, the system expands the time window to 8 seconds and reduces the motion smoothness weight to 5%. In the generated baseline confidence score, the grid density percentage is increased to 70% (63 points for an area with a density of 0.9 points / square meter), and the timeliness indicator is set to "STA-20," indicating that the static state is expected to last for 20 cycles. When the passenger gets up and leaves, the system detects a change in motion characteristics within 2 seconds and immediately switches back to dynamic tracking mode.

[0110] The system overlays grid layers vertically, with each layer spaced 0.5 meters apart. When a passenger boards, the system detects a change in altitude coordinates from +6.2 meters to +8.1 meters within 40 seconds. The baseline is updated using four-dimensional coordinates (with an added altitude dimension), and a vertical consistency factor (weighted at 15%) is incorporated into the confidence score. When an abnormal rate of altitude change is detected (e.g., a change of 0.8 meters within 1 second), the system activates a false alarm mechanism: it checks data from adjacent sensors to rule out positioning drift caused by elevator malfunctions.

[0111] In long-distance tracking of international transit corridors, the system handles the relay problem of reference points. Passengers walk from Terminal 3 to Terminal 2, a distance of 1.2 kilometers spanning 37 grid areas. The system establishes a linked storage structure for reference points: each new reference point records the coordinates of its predecessor node, and a position interpolation algorithm is initiated when the signal is interrupted. When passing through the low-signal area of ​​the underground corridor, the system generates temporary reference points based on a kinematic model and performs position correction after the signal is restored, with a maximum correction distance recorded as 3.2 meters.

[0112] The system simultaneously processes the trajectory data of 12 passengers, maintaining an independent baseline update queue for each target. When the trajectories of two passengers intersect at the lounge entrance, the system maintains target differentiation through gait feature analysis and baggage identification. The baseline data structure incorporates a target ID label, and the confidence score includes a target specificity factor (based on clothing color, baggage characteristics, etc.). Within a 5-minute observation period, the system successfully distinguished and continuously tracked all targets without any identity confusion incidents.

[0113] The system detected three groups of passenger families conversing at the exit, resulting in complex location overlap. The baseline update algorithm activated a high-density mode: temporarily reducing the grid scale to 1 meter and extending the time window to 6 seconds. By analyzing the relative positional patterns of the groups (e.g., children always staying within 1 meter of adults), the system correctly correlated the trajectories of each family member. Group relationship labels were added to the generated baselines, and a group consistency factor (weighted at 10%) was added to the confidence score.

[0114] As passengers walk along the curved glass curtain wall, the system uses a polar coordinate grid. Curvature parameters are introduced during reference point updates, and a turning rate of change index is added to the motion smoothness assessment. When a curvature change exceeding 15% is detected for three consecutive reference points, the system automatically increases the position sampling frequency by 50% to ensure smooth reconstruction of the curved trajectory. The final generated path accurately reproduces the passenger's movement along the 270-degree curved viewing walkway.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pedestrian trajectory tracking method based on artificial intelligence, characterized in that, Includes the following steps: Collect pedestrian movement feature data streams from multiple source sensors and extract key points of the initial trajectory as tracking reference points; The historical location data associated with the key points of the initial trajectory are divided into a set of high-confidence trajectory segments and a set of trajectory segments to be verified, and the set of trajectory segments to be verified is discarded. Receive a set of high-confidence trajectory segments, sort them in descending order by the number of trajectory points, select the top few segments as preferred trajectory segments, and calculate the proportion of preferred trajectory segments in the set of high-confidence trajectory segments; When the proportion is lower than the dynamic filtering threshold, a set of high-confidence trajectory segments is output as the initial tracking result; otherwise, the comparison trajectory data is output and a trajectory verification instruction is generated. In response to the trajectory verification command, the preferred trajectory segment is divided into spatial grids, and the trajectory optimization command is output in combination with the historical trajectory distribution characteristics. The comparison trajectory data includes grid division accuracy, historical data backtracking depth and optimization tolerance threshold. The tracking reference point is updated based on the trajectory optimization command, and the trajectory data of the secondary division is filtered and output. Obtain the proportion of preferred trajectory segments, the rate of change in the number of trajectory points, and the number of trajectory segments before and after instruction execution to generate a tracking efficiency evaluation coefficient; When the tracking efficiency evaluation coefficient is lower than the optimization threshold, the tracking parameter correction strategy is activated. The activation of the tracking parameter correction strategy is as follows: select the grid scale parameter in descending order of grid division score, and calculate the tracking efficiency evaluation coefficient using each scale parameter respectively. Select the grid scale parameter that causes the tracking efficiency evaluation coefficient to fall below the optimization threshold for the first time; Configure this grid scale parameter as the final tracking parameter.

2. The pedestrian trajectory tracking method based on artificial intelligence according to claim 1, characterized in that: Before extracting the key points of the initial trajectory, a spatiotemporal filtering algorithm is used to preprocess the sensor data stream; The pedestrian movement feature data stream is spatiotemporally encoded using a graph neural network, and the encoding results are fused with kinematic model parameters to generate a new trajectory prediction output.

3. The pedestrian trajectory tracking method based on artificial intelligence according to claim 1, characterized in that, The specific steps for segmenting historical location data include: Construct a trajectory confidence region centered on the key points of the trajectory; When the number of trajectory points within the confidence region is less than the preset trajectory density threshold, it is included in the set of trajectory segments to be verified. Conversely, they are classified into the set of high-confidence trajectory segments; When executing trajectory optimization instructions, if the growth rate of the number of trajectory points in the set of trajectory segments to be verified exceeds a preset percentage, it will be transferred to a set of high-confidence trajectory segments.

4. The pedestrian trajectory tracking method based on artificial intelligence according to claim 3, characterized in that, The steps for adjusting the dynamic filtering threshold include: Obtain the ratio of the number of trajectory segments to be verified to the number of high-confidence trajectory segments in the current scenario; When the ratio exceeds the upper limit threshold of trajectory density or falls below the lower limit threshold of trajectory density; Calculate the average trajectory point density for all trajectory segments; When the average trajectory point density is greater than the product of the upper limit threshold for trajectory density and the ratio, the upper limit threshold for trajectory density is increased. When the average trajectory point density is less than the product of the trajectory density lower limit threshold and the ratio, the trajectory density lower limit threshold is lowered.

5. The pedestrian trajectory tracking method based on artificial intelligence according to claim 4, characterized in that, The steps for selecting the preferred trajectory segment include: Retrieve the average effective trajectory length of historical tracking tasks; When the total number of high-confidence trajectory segments does not exceed the average effective trajectory length, the high-confidence trajectory segment set is output directly. Conversely, if the number of trajectory points is low, the trajectory segments with an equal length to the average effective trajectory length are selected as the preferred trajectory segments in descending order. If the continuous tracking time does not reach the preset period, retrieve the set of trajectory segments to be verified as supplementary trajectory data output.

6. The pedestrian trajectory tracking method based on artificial intelligence according to claim 5, characterized in that, The steps for generating trajectory verification instructions include: The conversion rate of unverified trajectory segments into high-confidence trajectory segments in historical tracking tasks was statistically analyzed. The number of trajectory points that are merged from the current set of high-confidence trajectory segments and the set of trajectory segments to be verified; When the calculated trajectory anomaly risk index exceeds the preset risk threshold, the set of high-confidence trajectory segments is directly output. Conversely, a trajectory verification command is generated.

7. The pedestrian trajectory tracking method based on artificial intelligence according to claim 6, characterized in that, The evaluation steps for spatial gridding include: The preferred trajectory segment is divided into multi-scale grids, and the difference between the maximum and minimum number of trajectory points in each grid cell is calculated. When the difference exceeds the spatial uniformity threshold, the grid cells are re-divided; Record the division scale parameters for qualified grid cells; A grid division score is generated by combining the distribution density and positional deviation of trajectory points in each grid cell; The scale parameter corresponding to the highest grid division score is selected as the trajectory optimization instruction.

8. The pedestrian trajectory tracking method based on artificial intelligence according to claim 1, characterized in that, The steps for calculating the tracking efficiency evaluation coefficient include: The rate of change of the number of preferred trajectory segments and trajectory points before and after the execution of the fusion instruction; Overlaying the trajectory anomaly risk index with preset weight parameters; When the total number of high-confidence trajectory segments does not exceed the average effective trajectory length, the normalization coefficient is used instead of the number of trajectory points for calculation.

9. The pedestrian trajectory tracking method based on artificial intelligence according to claim 7, characterized in that, The execution steps of the tracking parameter correction strategy include: Grid scale parameters are selected in descending order of grid division scores, and the tracking efficiency evaluation coefficients for each scale parameter are calculated. Select the grid scale parameter that causes the tracking efficiency evaluation coefficient to fall below the optimization threshold for the first time; Configure this grid scale parameter as the final tracking parameter.

10. The pedestrian trajectory tracking method based on artificial intelligence according to claim 9, characterized in that, The steps for updating the tracking benchmark include: Map the secondary segmented trajectory data to spatial grid cells; The coordinates of key points in the initial trajectory are corrected based on the density distribution of trajectory points within the grid cell. The updated tracking reference points are generated by integrating the correction results of continuous grid cells using a time sliding window.

Images