Embodied crowd simulation method based on imperfect stereovision

Abstract

The present application provides a kind of embodied perception crowd simulation method based on imperfect stereovision, to make the local behavior in crowd simulation more realistically reflect the perception constraint.The method is to set agent with different individual physiological attributes in 3D scene;For each agent, construct a differentiated embodied stereovision perception model;Through the model, binocular view is collected and converted into depth map, the depth map is divided into sectors and the depth information is aggregated, the sectorized free space representation is obtained;Based on the cognitive delay characteristics determined from individual physiological attributes, time filtering is performed on the free space representation to generate internal free space estimate;According to the internal free space estimate, the sector passability is judged, and the expected speed is generated based on the evaluation result of passable sector, and the non-collision position is updated.The present application unifies embodied difference, depth perception error and reaction time delay into behavior chain, so that the simulated movement is more naturally adapted to the change of visible space.

Description

Technical Field

[0001] This invention relates to the field of computer crowd simulation, and in particular to a method for embodied perception crowd simulation based on imperfect stereoscopic vision. Background Technology

[0002] Rapid urbanization and high-density public activities have made crowd simulation increasingly important in architectural design, evacuation assessment, and public safety management. Existing simulation systems can model the movement of large numbers of pedestrians in complex scenarios and are widely used in spatial layout analysis, capacity assessment, and risk prediction. For such applications, simulation results need to not only demonstrate reasonable pedestrian flow organization at the macro level but also reflect the local decision-making patterns of pedestrians in the real world at the micro level. Especially in environments with significant occlusion, limited field of vision, and dense interaction, pedestrian movement is not directly determined by complete and accurate environmental information but is simultaneously influenced by local visual input and the visible spatial structure.

[0003] Existing research has driven the development of crowd simulation technology from multiple perspectives. Classical geometric collision avoidance methods have achieved significant success in terms of real-time performance and stability; ORCA and its various extensions remain widely used representative baseline models in this field. Subsequently, researchers began to recognize the role of visual information in local navigation and proposed vision-based or synthetic vision-based methods that utilize the visible space from a first-person perspective to adjust the direction of movement. Subsequent methods have further optimized the coupling between visual control and local steering, while heuristic visual steering models have demonstrated the importance of visible space in pedestrian navigation.

[0004] In recent years, occlusion and local observability issues have received increasing attention. Related research indicates that limited field of view, occlusion, and the lack of local information significantly impact decision-making processes in crowded environments. However, while these methods address specific problems in local collision avoidance, visual guidance, or occlusion modeling, they typically rely on relatively idealized perception conditions, simplifying visual information into directly usable control signals. Therefore, local depth errors, field of view limitations, information loss due to occlusion, and the time delay between perception and action are difficult to integrate naturally into the behavior generation process. In scenarios such as blind spots, local congestion, and continuous two-way crowd flow, such simplification leads to overly idealized local interactions, necessitating additional rules to compensate for certain behavioral biases. Summary of the Invention

[0005] In view of the above situation, the main objective of this invention is to propose a method for embodied perception crowd simulation based on imperfect stereoscopic vision in order to solve the above-mentioned technical problems.

[0006] This invention proposes a method for embodied perception crowd simulation based on imperfect stereo vision, the method comprising the following steps: Step 1: Given a 3D scene, set up multiple intelligent agents in the 3D scene. The intelligent agents are configured with different individual physiological attributes. Step 2: Based on different individual physiological attributes, construct a differentiated embodied stereoscopic visual perception model for each intelligent agent; Step 3: Acquire binocular views using the embodied stereo vision perception model and convert the binocular views into depth maps; The depth map is divided into several sectors in the horizontal direction, and the depth information in each sector is aggregated to obtain a sectorized free space representation that represents the free space depth in each direction. Step 4: Based on the cognitive delay characteristics determined by the individual physiological attributes of the agent, perform temporal filtering on the sectorized free space representation to generate an internal free space estimate; Step 5: Determine the drivability of the current sector based on the internal free space estimation, and generate the expected speed for local turning based on the evaluation results of drivable sectors; perform collision-free position updates based on the expected speed to achieve crowd simulation.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. For the first time, embodied differences are explicitly integrated into the perception-behavior chain of population simulation, reproducing individual behavioral differences determined by physiological attributes.

[0008] Existing methods typically treat pedestrians as homogeneous individuals with uniform perceptual abilities, neglecting the fundamental impact of physiological factors such as height, body shape, and age on local environmental perception. This invention constructs a differentiated embodied stereoscopic vision perception model for each agent, mapping height to the virtual camera height and age and gender to perceptual parameters such as binocular baseline and perceptual field of view. This ensures that different agents "see" different free spaces within the same scene. This embodied modeling naturally reproduces behavioral variations in real-world populations, such as limited field of vision for shorter individuals and decreased stereoscopic perception accuracy in older individuals, without relying on external heterogeneous scripts or hard-coded rules.

[0009] 2. By modeling cognitive delay, the time mismatch between instantaneous perception and stable behavioral decision-making is resolved, resulting in a more human-like response pattern.

[0010] In reality, pedestrians' movement adjustments are not instantaneous responses to fleeting visual stimuli, but rather require a certain period of integration to form stable spatial judgments. Most existing methods directly use visual input as the control signal for the current moment, leading to overly sensitive behavioral responses and even high-frequency oscillations. This invention, based on the cognitive delay characteristics determined from individual physiological attributes, uses exponential time integration of the free space depth of each sector to effectively smooth local spatial estimation fluctuations caused by binocular stereo matching noise and transient occlusion. This processing not only suppresses directional jitter but also enables the agent to exhibit a highly consistent "observation-judgment-response" delay and smooth transition when encountering sudden visual obstruction or blind spots, just like real pedestrians.

[0011] 3. Replace traditional omnidirectional geometric collision avoidance with gap-driven steering, allowing local motion to directly respond to the visible spatial structure, rather than relying on invisible global information.

[0012] Traditional velocity-space optimization methods (such as ORCA) rely on complete and accurate neighborhood agent position and velocity information. In densely populated areas with severe occlusion and narrow field of vision, this can lead to overly aggressive avoidance or incorrect path selection due to information gaps. This invention starts with internal free space estimation, generating the desired velocity based solely on the accessibility and directional utility assessment of the currently visible sector. This desired velocity is then safely executed by a low-level collision-free solver. This hierarchical architecture ensures that the agent makes decisions only based on its own perceived space, thus enabling it to naturally and robustly exhibit local behaviors consistent with human spatial cognition, such as slowing down, waiting, and yielding, in scenarios like blind corner turns, local congestion, and dense two-way pedestrian traffic.

[0013] 4. A unified simulation framework has been established, which is flexible and compatible with classic local solvers and has good scalability and practicality.

[0014] This invention does not completely replace existing local obstacle avoidance algorithms. Instead, it uses embodied perception, cognitive delay, and gap-driven steering as upstream behavior generation modules, feeding the generated desired velocity into a mature solver (such as ORCA) for execution. This modular design allows the method to be seamlessly integrated into existing crowd simulation pipelines. It can independently improve the realism of behavior in perception-constrained scenarios without sacrificing the efficient obstacle avoidance performance of the underlying solver in open spaces, thus possessing high engineering application value.

[0015] In summary, this invention re-examines and models the core logic of "reacting to what one sees" in crowd simulation, starting from the source of perception. It provides a simulation solution with greater physiological rationality and behavioral expressiveness for application fields that have high requirements for behavioral realism, such as building design evaluation, evacuation simulation, and public safety management.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0017] Figure 1 This is a flowchart of the embodied perception crowd simulation method based on imperfect stereo vision proposed in this invention; Figure 2 This is an overview of the embodied perception and steering process of the present invention; Figure 3 This is a schematic diagram illustrating the instantiation of embodied perception parameters at the group level in the current implementation of the present invention; Figure 4 This is a qualitative comparison diagram between the present invention and the prior art in blind-angle scenarios; Figure 5 This is a comparison diagram of the qualitative results of the present invention and the prior art in a corridor scenario; Figure 6 This invention and the prior art BFR-DML-360-075-075 are set at the same reference time. Below is a diagram showing the synchronous longitudinal velocity distribution in the core area of ​​the channel; Figure 7 This is a comparison chart of the qualitative results of the present invention and the prior art in bottleneck scenarios; Figure 8 This is a qualitative comparison diagram of representative trajectory patterns of the present invention and existing technologies in a circular intersection scenario; Figure 9 This is a simulation diagram of a pedestrian scene at an urban intersection according to the present invention; Figure 10 This is a simulation diagram illustrating how to simultaneously eliminate population heterogeneity and cognitive delay in a bottleneck scenario. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0020] Example 1 Please see Figure 1This embodiment provides a method for embodied perception crowd simulation based on imperfect stereo vision, the method including the following steps: Step 1: Given a 3D scene, set up multiple intelligent agents in the 3D scene. The intelligent agents are configured with different individual physiological attributes. Step 2: Based on different individual physiological attributes, construct a differentiated embodied stereoscopic visual perception model for each intelligent agent; Step 3: Acquire binocular views using the embodied stereo vision perception model and convert the binocular views into depth maps; The depth map is divided into several sectors in the horizontal direction, and the depth information in each sector is aggregated to obtain a sectorized free space representation that represents the free space depth in each direction. Step 4: Based on the cognitive delay characteristics determined by the individual physiological attributes of the agent, perform temporal filtering on the sectorized free space representation to generate an internal free space estimate; Step 5: Determine the drivability of the current sector based on the internal free space estimation, and generate the expected speed for local turning based on the evaluation results of drivable sectors; perform collision-free position updates based on the expected speed to achieve crowd simulation.

[0021] Example 2 Please see Figure 2 This embodiment discloses the overall workflow of the present invention. Given a 3D scene and the current state of the agent, the system first constructs the agent's local perceptual conditions based on embodied parameters. These parameters include body size, binocular baseline, field of view, and cognitive delay characteristics. These parameters collectively determine how much local visual information the agent can acquire, and with what reliability such information can be converted into a usable local spatial representation.

[0022] During perception, pedestrians observe the local scene ahead through their left and right eye views and reconstruct a noisy local depth map. Due to limited field of view, unavoidable occlusion, and imperfections in binocular geometry, the resulting observations are often incomplete and noisy local descriptions of the environment. To transform these observations into a representation that can directly support action generation, this invention compresses the local depth map into a sector-based free-space description. This description characterizes the range of traversable space in different directions.

[0023] Subsequently, the method of this invention performs time integration on the sectorized depth observation results within a continuous time step. This process simulates the delay between instantaneous visual input and stable internal judgment. It smooths out short-term fluctuations in the perceived local space and generates a time-stable free-space representation for subsequent decision-making.

[0024] During motion generation, the method of this invention evaluates candidate traversal directions based on the current partitioned and sectorized free space representation, and selects a locally feasible direction based on its immediate visible structure. The selected direction is converted into a desired local velocity, then slightly smoothed, and passed to the ORCA solver. In this way, the agent can react to the currently visible local space while simultaneously utilizing the underlying collision avoidance backend.

[0025] Example 3 This embodiment discloses a method for configuring different individual physiological attributes of a given 3D scene and setting up multiple intelligent agents in the 3D scene.

[0026] An agent's perception of its environment is directly limited by its physiological structure. This invention places two virtual eyes in the local coordinate system of each agent. Each eye is implemented as a virtual camera, simulating binocular vision. The camera's height above the ground is determined by the agent's height. The horizontal distance between the two cameras is determined by the binocular baseline assigned to the agent. In the current implementation, the binocular baseline is sampled from gender-influenced physiological statistics and subsequently adjusted using age-based scaling. The perceptual field of view and cognitive delay constant are determined by a simple age-based prior distribution. Figure 3 This paper summarizes the specific implementation of this parameter across the entire population. Table 1 summarizes the key implementation parameters used in the embodied perception part, and the specific implementation methods are as follows: Table 1: Key implementation parameters used in this method.

[0027]

[0028] For each pedestrian agent i, the model obtains or generates the following individual parameters: height ,weight ,age ,gender Free walking speed Effective radius Binocular baseline Perceived field of view and cognitive delay constant .

[0029] If an external individual profile exists, basic individual parameters such as height, weight, age, and free speed are read from it; binocular baseline, perceived field of view, and cognitive delay constant can be further generated based on age, gender, and a preset mapping function. Specifically, height is sampled from a sex-related truncated normal distribution: ; in, The truncation operator for normal distribution sampling indicates sampling from a normal distribution within a given mean, standard deviation, and upper and lower bounds. These are the mean, standard deviation, minimum, and maximum heights corresponding to gender; Weight is indirectly derived from Body Mass Index (BMI). First, sampling... : ; in, Let be the body mass index of the i-th pedestrian. These are the mean and standard deviation of BMI corresponding to gender, as well as the minimum and maximum BMI values. Then calculate the weight using the following formula: ; Age is sampled from a truncated normal distribution: ; in, These are the mean, standard deviation, minimum age, and maximum age of the age distribution, respectively. Free walking speed is sampled from a truncated normal distribution: ; in, These represent the mean, standard deviation, minimum speed, and maximum speed of the free walking speed distribution, respectively. The effective radius of a pedestrian can be passed in from the scene, or it can be sampled within the range and then scaled. ; in, It is a very small positive number used to avoid radius degradation. These are the radius scaling factor and the original radius obtained from sampling, respectively.

[0030] Through the above process, each agent in the scenario acquires differentiated individual physiological attributes, providing basic parameters for building differentiated embodied perception models in subsequent steps.

[0031] Example 4 Please see Figure 3 Based on the individual physiological attributes generated in Example 3, this example constructs a unique embodied stereoscopic visual perception model for each intelligent agent. This model determines how the intelligent agent "sees" its surrounding environment, so that its perception ability matches its physical characteristics.

[0032] First, normalization is performed based on the agent's age. Define the normalized age variable: ; in, Let be the normalized age of the i-th pedestrian, with a value in the range [0, 1]. This is a truncation function used to restrict the input to the interval [0, 1]. These are the age-normalized young anchor points and old anchor points, respectively; Subsequently, normalized age was used Interpolation is performed between preset young and old parameter values ​​to generate three core perception parameters: The binocular baseline is determined by a combination of interocular distance statistics and age scaling. ; in, Interpupillary distance, in millimeters; The baseline for both eyes, These are the binocular baseline scaling factors for young and old individuals, respectively; The perceived field of view is obtained by interpolating the young field of view and the old field of view: ; in, To perceive the field of view, These are the young field of view and the old field of view, respectively. The cognitive delay constant is obtained by interpolating the youth delay and the old age delay: ; in, The cognitive delay constant, These are respectively youth-delayed and old-age-delayed; The above mapping enables different individuals to have different perceptual fields of view, depth accuracy, and reaction speeds in the same space.

[0033] Figure 3 The instantiation distribution of the above parameters at the group level is intuitively shown: the left figure shows the distribution of the binocular baseline with age, the middle figure shows the age-conditional prior of the perceived field of view, and the right figure shows the age-conditional prior of the cognitive delay constant.

[0034] Based on the above parameters, two virtual cameras are arranged in the local coordinate system of the intelligent agent to form the hardware layer of the embodied stereoscopic vision model. Let... For pedestrian location, For pedestrians, the local right direction is... Vertical direction Based on the eye height ratio, the centers of the left and right eye cameras are: ; ; in, Let the left eye camera center be the location of the i-th pedestrian. Let the right eye camera center be the i-th pedestrian. The eye-to-height ratio is used to determine the camera's mounting position in the body height direction. Let be the local rightward unit vector of the i-th pedestrian; The left and right viewpoint acquisition units move synchronously with the pedestrian's position and orientation. The distance between the left and right viewpoints is determined by the pedestrian's eye baseline, and the viewpoint height is determined by the ratio of the pedestrian's height to eye height. The viewpoint position of the binocular view of the virtual camera is determined by the center height and horizontal offset, so that the binocular view reflects the embodied differences caused by individual physiological attributes; the virtual camera moves synchronously with the position and orientation of the intelligent agent to obtain an embodied stereoscopic vision perception model.

[0035] Through the aforementioned differentiated parameter configurations, each agent acquires a unique stereoscopic perception capability; height affects visual field height, binocular baseline affects depth accuracy, and age affects both visual field range and reaction speed. This embodied difference lays the foundation for natural behavioral differences in subsequent steps.

[0036] Example 5 This embodiment discloses a specific implementation method for acquiring binocular views through an embodied stereo vision model and converting the binocular views into depth maps; dividing the depth maps into several sectors in the horizontal direction and aggregating the depth information in each sector to obtain a sectorized free space representation that represents the free space depth in each direction.

[0037] Based on the embodied stereo vision model constructed in Example 4, this example performs stereo perception and spatial representation transformation.

[0038] First, set the visual update interval. The simulation physical step size is Δt (e.g., 0.05s), and every [time / time]... Each simulation step performs one complete perception update. In this embodiment... =0.10s, meaning that visual information is updated every 0.10s, in order to simulate the cost of perception processing and reduce the computational load.

[0039] At the moment of visual update The agent's left and right virtual cameras simultaneously render the scene to obtain the left eye image. And right eye image And convert it to grayscale. and .

[0040] At the moment of visual update, the left and right eye cameras render images simultaneously: ; ; in, For a visual update moment, The left viewpoint is the camera center at the update time. The color image rendered for the left eye. The color image rendered for the right eye. Rendering operators that generate images based on camera position, pose, and scene. The camera center at the update time is the right viewpoint. The direction the pedestrian is facing or the camera's orientation; Then convert the left and right images to grayscale: ; in, This is a grayscale image of the left eye. This is the grayscale image of the right eye. For grayscale conversion operators, For left-eye view identification, This is the identifier for the right-eye view.

[0041] Next, the semi-global block matching (SGBM) algorithm was used to perform stereo matching on the left and right grayscale images. The SGBM parameters were set to: 64 disparity levels, block size 5, and fixed-point disparity multiplied by 16. The original disparity was then obtained. Afterwards, the actual parallax is: ; in, This represents the actual parallax. Calculate the horizontal pixel focal length based on the virtual camera parameters. Let the vertical field of view of the virtual camera be... If the image aspect ratio is W / H, then the horizontal field of view is: ; in, For horizontal field of view, Image width and height; The horizontal pixel focal length is: ; in, The horizontal pixel focal length; Based on the binocular geometry, the depth of each pixel is: ; in, For pixels The estimated depth at that location, in meters. This is the empirical focal length calibration factor.

[0042] The horizontal angle of the pixel column relative to the center line of sight is: ; in, For pixel columns The horizontal angle relative to the center line of sight of the camera; This results in a local depth map containing noise and errors.

[0043] Then, the dense depth map is converted into a sector-based free space representation, as follows: To suppress interference from both the sky and the ground, this invention analyzes only the scan band within a predefined vertical range in the image.

[0044] (1) Scan band filtering. Only pixels that fall within a predefined range in the vertical direction are processed, i.e., those whose row number y satisfies... ,in, For rounding, H is the image height. The percentage of the starting line of the scan band. The proportion of the end line of the scan band (in this embodiment) =0.35, =0.90), to filter out interference from the sky and ground areas.

[0045] (2) Horizontal sector division. Divide the horizontal range of the image into N sectors evenly. The sector index corresponding to pixel column x is: ; in, For pixel columns The corresponding sector index, This is a truncation function used to restrict the index to a valid range. The valid range for sector indexing is typically 0 to N-1.

[0046] (3) Robust aggregation of near / mid / far segments. Within each sector k, all samples that meet the effective depth condition are collected, i.e. ,in The minimum threshold for effective depth, For pixels The depth value is obtained by parallax conversion. The maximum threshold for effective depth. The minimum effective disparity threshold is used. Effective samples are divided into three ranges based on depth value: near range... medium range far range ,in, This is the upper bound of the near-distance segmentation. This is the upper bound of the mid-distance segment. This represents the upper bound of the effective depth. Let the three segments represent the depth as follows: ; in, These represent the near, middle, and far segments, respectively, and are used to replace the original depth of each sample in sector aggregation.

[0047] In the current implementation, valid depth samples are divided into near, medium, and far ranges, and the final sector depth is calculated based on their relative occupancy. This representation is more stable than using a single nearest depth sample and preserves the orientation structure required for local navigation. The resulting sector depth describes the amount of free space immediately available in each direction: ; ; ; in, For the effective sample set within a short distance, This is the effective sample set within the mid-range. It is a set of effective samples over a long distance.

[0048] Count the number of samples in each range , , If a sector contains valid samples, its free space depth estimate is a weighted average: ; in, For the free space depth estimation of the k-th sector, The sample sizes are for the near, middle, and far segments, respectively. If a sector has no valid samples, then we can set: ; in, This is the maximum effective depth; when there are no effective samples in a sector, it is used as a long-range estimate of the unobserved obstacle in that direction.

[0049] This aggregation method is more stable than taking the single nearest depth and can reduce the impact of single-pixel anomalies in binocular mismatches on pedestrian decision-making.

[0050] Finally, for agent i, at the current moment... A one-dimensional sectorized free space representation vector is obtained: this vector describes the range of traversable space in each direction, providing a direct basis for subsequent turning decisions.

[0051] Example 6 This embodiment discloses a specific implementation method for performing temporal filtering on sectorized free space representation based on the cognitive delay characteristics determined by the individual physiological attributes of the agent to generate the internal free space estimation part.

[0052] In the real world, pedestrians do not immediately adjust their movements based on visual observations; there is a delay between visual observation and stable behavioral responses. That is, the reaction to perceived changes in the scene is typically delayed, which can alter short-term collision avoidance behavior and accumulate into significant differences in crowd dynamics. Based on this observation, this invention introduces a cognitive delay mechanism on top of the aforementioned partitioned depth representation. Based on sectorized free-space representation, this embodiment simulates the cognitive delay process between instantaneous visual input and the formation of stable internal judgments by pedestrians, allowing different agents to produce differentiated response speeds due to their individual cognitive characteristics.

[0053] Visual perception does not need to be updated every physical frame, but rather at fixed visual update intervals. Updated. Simulation step size is... Then we can assume: ; Every Each simulation step calls the binocular perceptron once. This allows for the simulation of perception processing costs and reduces computational overhead.

[0054] For agent i, the cognitive delay constant determined in Example 4 is utilized. Calculate the exponential time integral coefficient: ; in, The integral coefficient over exponential time is... This is the cognitive delay time constant. This parameter is determined by the agent's physiological characteristics. In the current implementation, Depend on Figure 3 The age-based prior distribution shown is determined. Larger... This leads to a slower adaptation to newly observed local spaces. If it is extremely small or less than a preset threshold, then it can be set to... This indicates that the current visual input is immediately written into the internal state.

[0055] For each sector The current raw depth observation Convert to internal free space estimation The sector depth of the time integral is updated as follows: ; in, For the current visual update time, estimate the internal free space of the i-th pedestrian in the k-th sector. The original sector free space depth obtained from the current visual update. Estimate the internal free space reserved for the previous visual update moment; If this is the first time sensing, then initialize directly: ; in, This is an estimate of the internal free space after initialization during the first perception. This represents the initial free space depth of the original sector as perceived for the first time.

[0056] The effect of this processing is that short-term noise will not immediately cause a sudden change in direction; after the occlusion is restored, the agent's internal spatial judgment has a smooth transition, thus producing a reaction delay that is closer to that of a real pedestrian.

[0057] Figure 3: Instantiation of embodied perception parameters at the group level in the current implementation. Binocular baselines instantiated according to age among different agents (left). Age-conditional priors used to allocate the perceptual field of view (middle). Age-conditional priors used to allocate the cognitive delay constant (right).

[0058] Through the above processing, agents of different ages exhibit significant differences in their internal free-space judgments used for decision-making over time when faced with the same visual input. For example, when an obstacle suddenly appears in the field of vision, a younger agent's internal estimation almost immediately reflects spatial narrowing, while an older agent requires more time to fully incorporate this change into its internal representation. This temporal filtering, driven by cognitive delay characteristics, allows individual differences in reaction speed to naturally emerge in crowd simulations, reproducing behavioral differentiation caused by factors such as age without the need for additional artificial rules.

[0059] Example 7 This embodiment discloses a specific implementation method for determining the drivability of the current sector based on internal free space estimation, generating a desired velocity for local turning based on the evaluation results of drivable sectors, and performing collision-free position updates based on the desired velocity to achieve crowd simulation.

[0060] First, the passable sectors are determined by the free space depth of the sectors after cognitive delay. Then, the passable sectors are evaluated for utility. The sector with the highest score is selected as the local turning direction. Subsequently, the expected velocity is generated based on the effective free space in this direction. After light smoothing, the expected velocity is handed over to the ORCA collision-free local solver for execution.

[0061] Let the current simulation time be t for the i-th pedestrian. Its sector depth after embodied stereo perception and cognitive delay processing is represented as: ; Where N is the number of sectors divided within the horizontal field of view. Let be the available free space depth in the k-th sector after time integration. For simplicity, the pedestrian subscript i and time superscript t are omitted below, and it is denoted as . .

[0062] For each directional sector, first determine whether the delayed depth of that sector is sufficient to support pedestrian passage. A feasibility threshold determined by the pedestrian's physical radius is used. ; in, This is the radius scaling factor. For an additional safety margin, As a feasibility threshold, Let be the effective radius of the i-th pedestrian.

[0063] Traverse all N sectors. For sector k, if its internal free space estimate is greater than the passage threshold, i.e. ,in, If the above feasibility threshold is met, then the sector is marked as a passable sector. The set of passable sectors is denoted as: ; in, For the set of passable sectors, For sector index.

[0064] The above criteria mean that local turning is not simply about heading towards the target direction, but is primarily constrained by the current visible space and body dimensions. For pedestrians with larger body radii or greater safety margins, a larger amount of free space in front is required to consider a particular sector as passable.

[0065] For all sectors that meet the passability conditions, the utility score for that direction is further calculated. The sector utility in this invention consists of two parts: the first part is the alignment degree between the sector direction and the target direction, and the second part is the free space redundancy in that direction after exceeding the passability threshold. The utility function for sector k is defined as follows: ; in, The utility score for the k-th sector. This represents the non-negative alignment term between the sector direction and the target direction. The weight of the target direction alignment term. The weights for the free space depth term. To control the saturation rate of free space utility.

[0066] The meaning of this utility function is: the closer a certain direction is to the target direction, the better. The larger the value, the higher the directional alignment value; when the available depth in a certain direction... The greater the threshold, the higher the free space term becomes, but its gain gradually decreases through the exponential saturation function to prevent excessive depth from dominating the entire decision.

[0067] After obtaining the utility scores of each passable sector, select the sector with the highest score as the current local turning sector: ; in, For the current local steering sector, The candidate set consists of all passable sectors; Accordingly, the current local expected turning direction is: ; in, The current local desired direction of turning, This is the index of the optimal sector.

[0068] In other words, pedestrians do not directly use the complete neighborhood state to calculate the avoidance direction, but instead choose the most passable and efficient direction from the currently visible sector free space after cognitive delay stabilization.

[0069] After selecting a steering direction, the model scales the velocity based on the effective free space in that direction and generates the desired local velocity for the current step. The following explains how direction is further converted into velocity.

[0070] First, calculate the effective redundancy passage depth in the selected direction: ; in, The depth of the internal free space of the selected sector, The effective redundant passage depth in the selected direction, This is the feasibility threshold; Then map this redundant depth into a velocity scaling factor: ; in, This is the speed scaling factor. This is the distance scale used for deceleration control. To prevent pedestrians from coming to a complete stop in extremely narrow but still passable gaps, a minimum speed factor can be set, resulting in a scalar desired speed: ; in, For scalar expected velocity, The target speed is obtained by multiplying the individual's free walking speed by the scene speed. It is the minimum speed factor, used to ensure low-speed movement in narrow but passable gaps; This yields the desired velocity vector: ; in, Let be the desired velocity vector.

[0071] The meaning of this mechanism is: when there is sufficient redundant space in the selected direction, the agent moves at a speed close to that of free walking; when the space gradually narrows but is still passable, the agent automatically slows down, simulating the cautious behavior of a pedestrian in a narrow gap.

[0072] To avoid high-frequency directional jitter caused by discrete sector selection and depth noise, a light smoothing process is performed after generating the desired local velocity. This step does not change the basic logic of gap-driven steering; it is only used to limit overly abrupt steering and speed changes. Exponential smoothing is used. ; in, The expected velocity after smoothing at the current time step. The expected velocity after smoothing from the previous time step. To smoothly update the coefficients, these coefficients can be adjusted according to the simulation time step. and execution smoothing strength Defined as: ; When applying smoothing strength When the value is large, the smoothed speed approaches the current desired speed more quickly; when applying smoothing strength... When the size is smaller, the smoothing effect is stronger and the speed change is more slow.

[0073] The smoothed desired velocity is used as the preferred velocity input to the ORCA local solver. The ORCA solver collects the set of neighboring agents within the agent's perception range (determined by the current positions and radii of other agents) and static obstacle constraints in the scene, and calculates the collision-free actual velocity at the current moment based on geometric optimization: The smoothed expected velocity is not directly used as the final displacement velocity, but rather as the input to the underlying local collision-free solver. This process can be summarized as follows: ; in, Let i be the set of local neighbors related to the i-th pedestrian. For static obstacles or boundary constraints in the scene, To obtain the actual execution speed, For ORCA solvers.

[0074] Finally, update the agent's position according to the simulation time step: ; in, Let i be the position of the i-th pedestrian in the next simulation step. The actual execution speed after ORCA solution.

[0075] The aforementioned physical execution layer ensures that local movements do not result in collisions. This invention provides a desired velocity generated by embodied stereo perception, cognitive delay, and gap-driven steering, enabling the underlying collision-free actuator to operate under input conditions that more closely resemble real-world limited perception.

[0076] Through the sequential processing of the above five stages, the agent's local movements are always based on the limited information obtained from its embodied stereo perception and cognitive delay processing, rather than relying on the globally accurate neighborhood state. In scenarios with limited visibility, severe occlusion, or high-density interactions, this method can naturally reproduce behavioral differences caused by perceptual constraints, such as blind spot hesitation, gap selection preference, and reaction delay, thus providing a more realistic crowd simulation framework.

[0077] Example 8 This embodiment verifies the superiority of the present invention compared to other existing solutions, as detailed below.

[0078] Experimental Setup: All experiments were conducted in Unity using standard scene layouts. This invention considers four representative crowd interaction scenarios: blind corners, bottlenecks, corridors, and roundabouts. The blind corner scenario is implemented as two pedestrians orthogonally meeting at a corner, where the initial line of sight is completely obscured by the corner geometry. The bottleneck scenario follows the entrance setup from the experiment report on crowd gathering and passage before a narrow entrance, with pedestrians moving towards the narrow opening and forming local congestion and queuing near the bottleneck. The corridor scenario follows the laboratory setup of the BFR-DML two-way passage scenario, where pedestrians flow bidirectionally through a narrow passage. The roundabout crossing scenario follows the classic symmetric benchmark of the synthetic vision turning method, with pedestrians initialized in a roundabout area, and each agent assigned a point on the opposite side of the diagonal as the target, thus forming the most conflict-prone open area crossing task. All simulations were conducted within these standard benchmark geometries. For visualization purposes only, the generated trajectories were replayed in a uniform city street view to ensure consistency of scene background between the illustrations and videos. This replay stage is purely for demonstration purposes and does not affect the simulation dynamics, control logic, or quantitative measurement results. All numerical evaluations reported in this paper were calculated in the original simulation coordinate system, not in the remapped playback scene.

[0079] This invention compares three methods: the ORCA method, the synthetic visual steering method, and the method of this invention. The ORCA method serves as the geometric collision avoidance benchmark. The synthetic visual steering method serves as the visual heuristic steering benchmark. The method proposed in this invention is an embodied perception framework that integrates imperfect perception, cognitive delay, and gap-based local steering strategies. All three methods employ the same scene geometry, generation configuration, target specifications, and simulation time step. For scenarios requiring intermediate pathpoint support, all methods use the same waypoint structure to minimize the impact of external path design. This invention employs evaluation criteria tailored to different scenarios. For blind-angle scenarios, this invention examines occlusion recovery and yielding behavior through qualitative analysis; for bottleneck scenarios, this invention analyzes local traffic efficiency and congestion near entrances; for corridor scenarios, this invention performs qualitative evaluation by maintaining bidirectional order and motion continuity; for circular intersection scenarios, this invention reports the maximum crossing time and the proportion of slow-moving pedestrians.

[0080] Blind Spots: Blind spot scenarios directly reveal the behaviors of occlusion recovery and yielding under limited visibility conditions. Two pedestrians approach the same intersection from opposite sides of a building corner. Initially, the building completely obstructs their view. As they gradually approach each other, visibility gradually recovers, and localized yielding behaviors begin to emerge. Figure 9 The scenario is illustrated. The top row presents a real-world reference case from the Pedestrian Dynamics Data Archive dataset. It provides a behavioral reference for corner encounters caused by occlusion in real pedestrian movement. The following rows show representative simulation sequences of ORCA (second row), the synthetic visual turning method (third row), and the method proposed in this invention (bottom row). The three methods rely on different forms of local information. The ORCA method, which mainly relies on geometric states for collision avoidance, is prone to hesitation or local stagnation under symmetrical approach conditions; the synthetic visual turning method can generate a turning response relatively quickly after visibility is restored, but its behavior is more dependent on instantaneous visual changes. The method of this invention further incorporates imperfect depth perception and cognitive delays, so local behavior depends on noisy, constrained, and temporally integrated representations in the visible space. Figure 9 As shown, these differences lead to distinctly different interaction patterns near corners. Under symmetrical approach conditions, ORCA is more prone to mutual hesitation or low-speed stagnation. Synthetic visual steering methods can initiate steering relatively quickly after visual field recovery, but their response remains tightly coupled with immediate visual changes. The method of this invention updates local depth observations after visual field recovery, constructs a more stable internal representation through cognitive delay, and then uses gap-driven steering to generate the desired speed. This enables pedestrians to achieve a more coherent yielding and overtaking process near corners.

[0081] Bottlenecks and Corridors: This invention further evaluates the proposed method in two classic scenarios—corridors and bottlenecks—to examine its behavior under continuous bidirectional interaction and locally congested flow fields. In these cases, the focus shifts from behavior triggering after occlusion recovery to the stability of the method under continuous bidirectional interaction and its robustness in congested channels. The corridor scenario adopts a standard bidirectional straight corridor setting commonly used in pedestrian flow experiments. The Unity (a 3D rendering engine) implementation of this invention is instantiated based on the BFR-DML settings in previous research, employing a balanced flow and dynamic multi-lane configuration. Figure 5 The qualitative results comparison chart for the corridor scene shows the representative results of this scene. Figure 5 In the diagram, ORCA (left), the method of this invention (middle), and the synthetic visual steering method (right) are shown. As can be seen, the ORCA method mainly relies on geometric collision avoidance constraints, which easily leads to strong local avoidance and formation disturbances in two-way traffic. The synthetic visual steering method can produce a certain visual response, but the local motion organization is still relatively scattered. The method of this invention, through embodied binocular perception, passable space judgment, and cognitive delay modeling, enables pedestrians to form a more continuous traffic organization in two-way channels, reducing local stagnation and frequent turning, and exhibiting a more stable dynamic multi-lane traffic effect.

[0082] Figure 6 The instantaneous longitudinal velocity distribution within the core area of ​​the corridor was compared at the same synchronization reference time (t = 23.5s). Adjacent Voronoi regions (filled with the same color) indicate a more coherent local opposing flow organization. The number (n) in each subfigure represents the instantaneous number of pedestrians within the core measurement area of ​​the corridor. As shown in the figure, compared with other methods, the method of this invention produces a clearer velocity distribution with less fragmentation, forming clearer positive and negative velocity regions, reducing isolated low-velocity units and directionally mixed regions, indicating that this invention can maintain good longitudinal velocity consistency and channel organization under continuous bidirectional flow conditions. All three methods support bidirectional passage, but differ in local organization. ORCA continues to exhibit strong geometrically driven collision avoidance capabilities. The synthetic visual steering method maintains relatively smooth local avoidance during continuous interaction. The method of this invention also relies on local perception, but the resulting bidirectional flow remains continuous and stable, without continuous hesitation or frequent yaw, and exhibits a more coherent opposing flow structure in the core area of ​​the corridor.

[0083] The bottleneck scenario was constructed following the entrance bottleneck setting in previous empirical studies. Pedestrians move towards a narrow passage, creating localized congestion and queue-like formations near the entrance. This scenario aims to test whether the method remains effective when interactions are compressed into a limited area. Figure 7Typical results are shown in the diagram for bottleneck scenarios, including ORCA (left), the method of this invention (middle), and the synthetic visual steering method (right). All three methods can create a passage at the entrance, but the final crowd arrangement differs. ORCA tends to make conservative local adjustments near narrow passages, which may result in lower traffic efficiency. While the synthetic visual steering method improves local responsiveness, its movement is still directly affected by instantaneous visual changes. The method of this invention supports a more orderly and passable local organization near the entrance, while maintaining a more continuous yielding and passage process without significant persistent congestion. These results demonstrate that under the proposed model, stable local organization and continuous passage capacity near the entrance enable pedestrians to form a more orderly arrangement and passage process at bottlenecks, reducing persistent congestion and improving traffic stability in narrow passages.

[0084] Roundabouts: Table 2 summarizes the quantitative results for roundabout crossing scenarios. All three methods successfully completed the task. In this scenario, the quantitative differences between the methods were moderate, but the method of this invention achieved the best average results in both metrics: the shortest maximum crossing time and the lowest proportion of slow-moving pedestrians. Figure 8 A qualitative comparison of representative trajectory patterns is presented, including ORCA (left), the proposed method (middle), and the synthetic visual steering method (right). The ORCA method typically produces relatively direct radial intersection trajectories with a wide distribution of interaction areas; the synthetic visual steering method can induce some detours, but the local interaction areas are relatively concentrated. The proposed method influences local steering decisions through imperfect binocular perception and delayed perception states, resulting in a more coherent exchange structure and smoother interaction patterns. This demonstrates that the proposed method can maintain effective local avoidance and passage capabilities in open intersection scenarios.

[0085] Table 2: Comparison of results for three methods (ORCA, synthetic visual steering, and the method of this invention) under 10 random initialization conditions for circular intersection crossing scenarios.

[0086]

[0087] In Table 2, It indicates a decrease.

[0088] As shown in Table 2, all methods successfully completed all tests. The shorter the maximum travel time and the lower the proportion of slow travelers, the better the performance.

[0089] This invention also constructs a realistic scene of pedestrians at an urban intersection, including both pedestrians crossing in a straight line and those turning to cross. At each intersection, pedestrians heading to different crosswalks form independent waiting groups near their respective crossing entrances. After crossing, pedestrians further disperse along adjacent sidewalks, resulting in rich interactions in the corner areas and the center of the intersection. Figure 4 As shown, the scenario involves over 500 participants and depicts pedestrians at a city intersection, including those crossing in a straight line and those turning. Pedestrians wait at their respective crossing points, and after crossing the intersection, disperse along adjacent sidewalks, demonstrating that the method of this invention is also applicable to large-scale crowd simulations.

[0090] To examine the respective roles of embodied heterogeneity and cognitive delay within the framework of this invention, an ablation analysis was performed in a bottleneck scenario. The complete model was compared with two simplified variants: (i) no delay (w / oDelay), which retains heterogeneous embodied features but removes cognitive delay; and (ii) no heterogeneity (w / o Hetero), which retains cognitive delay but replaces the heterogeneous embodied features of all pedestrians with shared features. Apart from the above modifications, the scene geometry, initial layout, target settings, and simulation time step remained consistent.

[0091] This invention selects a bottleneck scenario for ablation research because the continuously shrinking local space in this scenario amplifies short-range interactions and makes differences in local organizational structures more easily observed. Table 3 reports three quantitative indicators from multiple runs: (i) throughput, which measures the bottleneck passage efficiency; (ii) dwell time near the bottleneck, which measures the duration of low-speed congestion in the bottleneck area; and (iii) bottleneck throughput rate, which measures the proportion of pedestrians who successfully pass through the bottleneck.

[0092] Quantitative results show that removing any component significantly degrades system performance. The complete model achieves not only the highest throughput and bottleneck throughput but also the lowest dwell time near the bottleneck. Removing cognitive delay moderately reduces passage efficiency and induces higher localized congestion. Removing embodied heterogeneity further deteriorates performance, leading to more persistent congestion near the bottleneck and a decrease in the overall proportion of pedestrians passing through the bottleneck. These results indicate that, under bottleneck constraints, embodied heterogeneity and cognitive delay together contribute to a more stable and less compressed passage process.

[0093] Table 3: Ablation results after multiple runs in bottleneck scenarios.

[0094]

[0095] In Table 3, Indicates a decrease. This indicates an increase. Higher throughput and bottleneck throughput are better, and lower dwell time near the bottleneck is better.

[0096] Figure 10 A qualitative visualization of the two settings for joint removal was presented, simultaneously removing embodied heterogeneity and cognitive delay. Compared to the full model, the joint removal experiment resulted in a denser and more compact congestion cluster in front of the bottleneck. In the full model, pedestrians maintained small interpersonal distances and formed a less compressed local arrangement before reaching the exit. This qualitative trend is consistent with the results of the quantitative removal experiment, indicating that the perception of embodied heterogeneity and cognitive delay together contribute to maintaining a more dispersed and reliable local organizational structure near the bottleneck.

[0097] Figure 10 In bottleneck scenarios, this invention simultaneously eliminates crowd heterogeneity and cognitive delay. The leftmost sub-image shows a real-world laboratory entrance experiment as a behavioral reference. The middle image shows the default settings of the method, which include heterogeneous embodied characteristics and pedestrian-specific cognitive delay. The rightmost image removes both factors. All 22 pedestrians share the same characteristics, and cognitive delay is disabled. The red circles highlight the qualitative differences in local organizational structure near the bottleneck; the method of this invention maintains a small interpersonal gap before the exit, while the setting that removes these two factors evolves into a denser, more compact congestion cluster.

[0098] In summary, this invention models the local perception process based on imperfect stereoscopic vision perception. The framework maps height, body shape, and binocular baselines to local depth perception conditions, approximates cognitive delay through time integration, utilizes gap-driven steering to generate desired velocities, and delegates collision-free updates to the underlying actuator. Experimental results show that this framework can generate local interaction patterns consistent with spatial changes at blind spots, while maintaining stability and usability in scenarios such as corridors and bottlenecks. These results demonstrate that explicitly incorporating microscopic perception constraints into the behavior generation layer provides a modeling method for local interactions in crowd simulations that more closely approximates real-world visual conditions.

[0099] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for embodied perception crowd simulation based on imperfect stereo vision, characterized in that, The method includes the following steps: Step 1: Given a 3D scene, set up multiple intelligent agents in the 3D scene. The intelligent agents are configured with different individual physiological attributes. Step 2: Based on different individual physiological attributes, construct a differentiated embodied stereoscopic visual perception model for each intelligent agent; Step 3: Acquire binocular views using the embodied stereo vision perception model and convert the binocular views into depth maps; The depth map is divided into several sectors in the horizontal direction, and the depth information in each sector is aggregated to obtain a sectorized free space representation that represents the free space depth in each direction. Step 4: Based on the cognitive delay characteristics determined by the individual physiological attributes of the agent, perform temporal filtering on the sectorized free space representation to generate an internal free space estimate; Step 5: Determine the drivability of the current sector based on the internal free space estimation, and generate the expected speed for local turning based on the evaluation results of drivable sectors; Collision-free position updates are performed based on the desired velocity to achieve crowd simulation.

2. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 1, characterized in that, In step 1, individual physiological attributes include height, weight, age, and sex. The process of generating individual physiological attributes specifically includes the following steps: The age is normalized to the [0,1] interval, and interpolation is performed between preset young parameter values ​​and old parameter values ​​based on the normalized age to generate binocular baselines, perceptual field of view and cognitive delay constant, so that different intelligent agents can obtain differentiated stereo perception capabilities.

3. The embodied perception crowd simulation method based on imperfect stereoscopic vision according to claim 2, characterized in that, In step 2, a differentiated embodied stereoscopic visual perception model is constructed for each agent based on different individual physiological attributes, specifically including the following steps: The center height of the virtual camera is determined based on the height of the intelligent agent and the preset eye height ratio; Determine the horizontal offset of the two virtual cameras based on the baseline of both eyes and the current orientation of the agent; The viewpoint position of the binocular view of the virtual camera is determined by the center height and horizontal offset, so that the binocular view reflects the embodied differences caused by individual physiological attributes, thus obtaining an embodied stereoscopic visual perception model.

4. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 3, characterized in that, In step 3, the binocular view is converted into a depth map, which specifically includes the following steps: Perform semi-global block matching on the left and right binocular views to obtain disparity maps; Based on the virtual camera parameters in the embodied stereo vision perception model, the disparity map is restored to a depth map.

5. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 4, characterized in that, In step 3, the depth map is divided into several sectors along the horizontal direction, and the depth information within each sector is aggregated to obtain a sectorized free space representation that characterizes the free space depth in each direction. This specifically includes the following steps: The depth map is divided into several sectors in the horizontal direction, and the scanning band range is defined in the vertical direction of the image. Within each sector, depth samples that fall within the scanning zone and meet the effective depth range are divided into three ranges—near, middle, and far—based on their depth values ​​to obtain the representative depth of the samples. Based on the number of samples and the representative depth of the samples in each range, the free space depth of the corresponding sector is calculated by weighting, and the sectorized free space representation that represents the free space depth in each direction is obtained.

6. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 5, characterized in that, In step 4, based on the cognitive delay characteristics determined by the individual physiological attributes of the agent, temporal filtering is performed on the sectorized free space representation to generate an internal free space estimate. Specifically, this includes the following steps: The binocular view acquisition and the generation of sectorized free space representation are triggered at a preset visual update interval; Determining the cognitive delay constant from individual physiological attributes; For each sector, the exponential time integral coefficient is calculated using the cognitive delay constant; By using the exponential time integral coefficient, the sectorized free space representation at the current visual update moment is weighted and summed with the internal free space estimate at the previous moment to obtain the updated internal free space estimate.

7. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 6, characterized in that, In step 5, the accessibility of the sector is determined based on the internal free space estimation, which specifically includes the following steps: The passage depth threshold is determined based on the physical radius of the intelligent agent and the preset safety margin; Sectors whose internal free space estimate is greater than the passage depth threshold are classified as passable sectors.

8. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 7, characterized in that, In step 5, the desired speed for local steering is generated based on the evaluation results of the traversable sector, specifically including the following steps: Calculate the utility score for each passable sector; the utility score is determined by the alignment between the sector direction and the target direction, as well as the redundancy of the free space estimate within the sector that exceeds the passability depth threshold. Select the sector with the highest utility score as the desired turning direction; A speed scaling factor is generated based on the redundancy of the selected sector, and a scalar expected speed is obtained based on the speed scaling factor and a preset free-walking speed. The desired scalar velocity is combined with the desired steering direction to generate the desired velocity vector.

9. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 8, characterized in that, After generating the desired velocity vector and before performing collision-free position updates, the process also includes: performing exponential smoothing on the desired velocity vector to suppress directional jitter caused by sector selection and depth noise.

10. The embodied perception crowd simulation method based on imperfect stereo vision according to claim 9, characterized in that, In step 5, collision-free position updates are performed based on the desired velocity. Specifically, the desired velocity is used as the preferred velocity input to the ORCA solver. The ORCA solver calculates the actual collision-free velocity based on the constraints of neighboring agents and static obstacles, and updates the agent's position.

Images