Non-motor vehicle trajectory reconstruction method in motor-vehicle-shared space

Abstract

The application relates to a non-motor vehicle track reconstruction method in a motor-non-motor shared space, and the method steps comprise the following: extracting an initial track through a CV algorithm; adopting an LSTM to predict and complete a track missing due to occlusion, and matching the track before and after occlusion through a Hungarian algorithm to complete a complete track; adopting a three-layer cyclic social force particle filtering algorithm constructed by integrating a social force model representing non-motor vehicle driving behavior into a particle filtering framework to denoise the completed complete track to obtain a reconstructed track. Compared with the prior art, the application can effectively improve the performance of a video detector in acquiring high-precision track data, provide more fine perception data for a holographic intersection and a traffic control system, and has great application value in the industry.

Description

Technical Field

[0001] This invention relates to the field of holographic intersections and active traffic control, and in particular to a high-precision trajectory reconstruction method that focuses on computer vision algorithms to extract trajectory missing and severe noise problems in roadside videos of non-motorized vehicles, and integrates a micro-interactive behavior model into a filtering algorithm. Background Technology

[0002] The development of computer vision (CV) algorithms has greatly promoted the application of video detectors in the transportation industry, especially in acquiring high-precision vehicle trajectories from video. This plays a crucial role in the development of fields such as holographic intersections, active traffic control, and autonomous driving. Existing CV algorithms can effectively capture the trajectories of large targets such as motor vehicles with large external contours and small morphological changes. However, significant technical obstacles remain for the trajectories of small targets such as non-motorized vehicles and pedestrians with smaller shapes and greater morphological changes. In particular, the trajectory accuracy is greatly reduced in shared spaces with high-density mixed traffic flow of motor vehicles and non-motorized vehicles.

[0003] In shared spaces, the mixing of motorized and non-motorized traffic is severe, with mutual interference between vehicles. Roadside traffic videos exhibit significant occlusion issues, leading to severe trajectory loss and ID inconsistencies when directly using CV (Cartography) algorithms to extract vehicle trajectories. Furthermore, sub-pixel-level jumps in the boundaries of detected objects cause significant deviations between the detected target center point coordinates and actual values, resulting in severe noise in the output non-motorized vehicle trajectories. Existing trajectory extraction algorithms focus on extracting and denoising motorized vehicle trajectories, with few methods for extracting and reconstructing non-motorized vehicle trajectories. However, significant morphological and motion characteristic differences exist between non-motorized and motorized vehicles. CV-outputted motorized vehicle trajectories rarely suffer from trajectory loss due to occlusion, exhibiting only localized trajectory noise. Therefore, existing motorized vehicle trajectory reconstruction algorithms are difficult to directly apply to reconstructing non-motorized vehicle trajectories. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a high-precision non-motorized vehicle trajectory reconstruction method in a shared space between motorized and non-motorized vehicles.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for reconstructing the trajectory of non-motorized vehicles in a shared space for motorized and non-motorized vehicles, the method comprising the following steps:

[0007] The initial trajectory is extracted using the CV algorithm;

[0008] LSTM prediction is used to complete the trajectory missing due to occlusion in the initial trajectory, and the Hungarian algorithm is used to match the trajectories before and after occlusion to complete the complete trajectory.

[0009] A three-layer cyclic social force particle filter algorithm, which incorporates a social force model representing non-motorized vehicle driving behavior into a particle filter framework, is used to denoise the completed trajectory and obtain the reconstructed trajectory.

[0010] Furthermore, the missing trajectory is completed by constructing a prediction model using a two-layer LSTM structure. The input of the prediction model is the trajectory acceleration sequence before occlusion, and the output of the prediction model is the missing trajectory caused by occlusion. The two layers of LSTM are used to predict longitudinal acceleration and lateral acceleration, respectively.

[0011] Furthermore, the goal of matching trajectories before and after occlusion is to obtain a complete trajectory. This considers the consistency characteristics of the trajectories before and after occlusion, and also considers fusing the temporal and spatial feature information of the trajectories before and after occlusion. The objective function for matching is:

[0012]

[0013] In the formula, x k,h C is the decision value for matching the k-th trajectory in the predicted trajectory set of the occluded region with the h-th trajectory in the trajectory set after occlusion; k,h For matching costs;

[0014] The matching cost C k,h The equation is constructed using the Euclidean distance method and solved using the Hungarian algorithm, as shown below.

[0015]

[0016] In the formula, It is the average value of the global time frame corresponding to the last second of the predicted trajectory set of the occluded area and the average value of the trajectory sequence; It is the average value of the time frame corresponding to the first second in the trajectory set after occlusion and the average value of the trajectory sequence.

[0017] Furthermore, based on the matching scheme obtained from the Hungarian algorithm, the predicted trajectory set and the trajectory set after the occlusion area are concatenated, and the integrity of the trajectory is determined to obtain the complete trajectory; the rules for the integrity determination are as follows:

[0018] If the last frame of the predicted trajectory set is equal to the first frame of the trajectory set after the occlusion area, then it is a complete trajectory.

[0019] If the last frame of the predicted trajectory set is smaller than the first frame of the trajectory set after the occlusion area, then the missing frame is interpolated using the linear interpolation method.

[0020] If the last frame of the predicted trajectory set is larger than the first frame of the trajectory set after the occlusion area, then delete the trajectory corresponding to the extra frame in the predicted trajectory set.

[0021] Furthermore, the aforementioned cyclic social force particle filter algorithm

[0022] This paper integrates the interactive behavior model in microscopic traffic flow with the particle filter algorithm. The particle filter algorithm, based on the unrestricted noise distribution assumption, serves as the trajectory fusion unit. The social force model is used as the state update equation in the particle filter, and the CV trajectory is used as the input value for the observation update equation. It contains three layers of loops, and the specific functions and implementations of each layer are as follows:

[0023] The outer loop performs multiple rounds of noise reduction processing on all noise trajectories. The new round will execute the noise reduction algorithm based on the results of the previous round. After the multiple rounds of loops are completed, the most stable result of the iteration is selected as the final result.

[0024] The middle loop performs noise reduction processing on all vehicles and replaces the corresponding vehicle ID trajectory in the environmental traffic flow with the noise reduction results;

[0025] The inner loop fuses the CV trajectory and the social force-generated trajectory frame by frame. At the end of one loop, the state equation module updates the vehicle motion state according to the social force model and reads the original noise trajectory state value.

[0026] Furthermore, the described cyclic social force particle filter algorithm uses a social force model to measure and update the CV trajectory in the fusion algorithm, and the state update equation is shown below:

[0027]

[0028] In the formula, Represents self-driving force; ∑ j(j≠b) F b，j (t) represents the interaction forces between non-motorized vehicles; ∑ W F bW (t) represents the boundary force; ∑ car F bc (t) represents the force exerted by the motor vehicle; ξ is the random error; the parameters of the social force model are calibrated using a genetic algorithm.

[0029] Furthermore, the expression for the self-driving force is shown below:

[0030]

[0031] In the formula, v is the expected speed of vehicle b at time t; b (t) The current speed of vehicle b at time t; τ b The average time required to change the relaxation time from the current speed to the periodic speed; a b Represents the expected acceleration; m b It is the mass of vehicle b.

[0032] Furthermore, the expression for the interaction force between non-motorized vehicles is shown in the following formula:

[0033]

[0034] In the formula, It is the repulsive force between adjacent non-motorized vehicles j at time t. For overtaking force, ω bj For the same directional dependency weights;

[0035] The expression for the repulsive force of non-motorized vehicle j on non-motorized vehicle b is as follows:

[0036]

[0037] In the formula, The minor semi-axis of the ellipse drawn with the location of non-motorized vehicle b as its focus; (v b -v j ) is the speed of non-motorized vehicle b relative to j; U3 and R3 represent the scaling factor and the distance sensitivity coefficient, respectively; r bj It is the vector from non-motorized vehicle j to b; Δt is the time step; n bj It is the direction vector from non-motorized vehicle j to b;

[0038] The overtaking force is shown in the following formula:

[0039]

[0040] In the formula, k is the proportionality coefficient; It is perpendicular to And with (v) b -v j Unit vectors with smaller included angles;

[0041] The same direction-dependent weight ω bj The expression is as follows:

[0042]

[0043] In the formula, λ b The range of variation is 0 << λ b <<1; It is the angle between the position vector of non-motorized vehicle b relative to j and the direction of travel of non-motorized vehicle b.

[0044] Furthermore, the magnitude of the boundary force is related to the distance between the non-motorized vehicle and the lane boundary, as shown in the following expression:

[0045]

[0046] In the formula, U4 is the proportionality coefficient of the boundary repulsion force; r bw R4 is the distance between the non-motorized vehicle position and the lane boundary; R4 is the sensitivity coefficient of the boundary repulsion force to distance; n bw It refers to the unit vector pointing from the center of the non-motorized vehicle to the position on the lane closest to the non-motorized vehicle.

[0047] Furthermore, the expression for the force exerted by the motor vehicle is shown in the following formula:

[0048]

[0049] In the formula, U6 and R6 are the parameters to be estimated; r bc It is the distance between the motor vehicle and the target vehicle; ||v c || refers to the speed of the motor vehicle, n bc A unit vector representing the interaction force between motor vehicles and non-motor vehicles;

[0050] If there is a motor vehicle in front of the non-motorized vehicle, the non-motorized vehicle will experience a force that slows it down:

[0051]

[0052] S = s0 + v b Δt

[0053] In the formula, b c s0 is the comfortable deceleration speed for non-motorized vehicles; v is the safe distance for non-motorized vehicles; b Δt is the target non-motorized vehicle speed; d is the reaction time; bc This refers to the distance between motor vehicles and non-motor vehicles, n v It is the unit vector of deceleration affected by motor vehicles.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] This invention proposes a second-order trajectory reconstruction method considering vehicle interaction behavior: The first stage addresses the issues of missing non-motorized vehicle trajectories and ID fluctuations by integrating the prediction function of the Long Short-Term Memory (LSTM) model with the matching function of the Hungarian algorithm to form an algorithm for trajectory completion in occluded areas and trajectory matching after occlusion. The second stage focuses on handling the noise problem of non-motorized vehicle trajectories by fusing a social force model representing the micro-behavior of mixed traffic flow with a particle filter algorithm to construct a three-layer cyclic filtering algorithm. This method ensures that the reconstructed trajectory retains the morphological characteristics of the CV output trajectory while satisfying the constraints of vehicle micro-driving behavior under mixed traffic flow. Furthermore, the outer loop employs iterative convergence and multiple iterations to find the optimal value, effectively avoiding random errors. This invention can effectively improve the performance of video detectors in acquiring high-precision trajectory data, providing more refined perception data for holographic intersections and traffic control systems, and has significant application value in industry. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the process for high-precision non-motorized vehicle trajectory reconstruction in a shared space for motorized and non-motorized vehicles according to the present invention.

[0057] Figure 2 This is a schematic diagram of the high-precision trajectory extraction flow method based on roadside view video;

[0058] Figure 3 This is a schematic diagram of the two-dimensional trajectory reconstruction algorithm of the present invention;

[0059] Figure 4 This is a schematic diagram of the process of predicting trajectories using LSTM in this invention;

[0060] Figure 5 This is a schematic diagram of the trajectory reconstruction algorithm that integrates social forces and particle filtering according to the present invention. Detailed Implementation

[0061] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0062] Example 1

[0063] To address the issues of trajectory loss, ID inconsistencies, and severe noise in trajectories extracted by CV algorithms in complex traffic environments with shared traffic spaces, this invention integrates theories from multiple disciplines, including computer vision, microscopic traffic flow models, and automated control, to propose a high-precision trajectory reconstruction method for roadside traffic videos in shared spaces. This method aims to acquire high-precision trajectory data for all objects within the shared space, establishing a digital foundation for engineering practices such as autonomous driving, holographic intersections, and active traffic management, ultimately improving traffic efficiency and driving safety. The method mainly consists of three modules: video preprocessing, object detection and tracking, and trajectory reconstruction. The overall technical framework is as follows: Figure 1-2 As shown, the first two modules already incorporate technologies from the field of computer vision.

[0064] To address the issues of missing local trajectories and severe noise in the initial trajectories output during the target detection and tracking phase, a second-order trajectory reconstruction method is proposed. In the first stage, the prediction function of the long-term memory model and the matching function of the Hungarian algorithm are integrated to complete the missing trajectories in occluded areas and match them with the trajectories of occluded jump IDs. In the second stage, a filtering algorithm considering micro-interaction behavior is created by fusing the social force model and particle filtering algorithm from the theory of micro-traffic flow micro-behavior. The complete trajectories of motor vehicles and non-motor vehicles directly output by the CV and the trajectory completed in the first stage are used as background traffic flow trajectories. A state update equation driven by environmental perception and the social force model is constructed, and the updated trajectory is fused with the original trajectory output by the CV frame by frame to improve trajectory accuracy.

[0065] Specific implementation steps:

[0066] This invention relates to a video preprocessing, target recognition, and tracking module that utilizes existing mature computer vision (CV) technology. The key focus of the invention is trajectory reconstruction. The implementation process of this module is as follows: Figure 3 As shown, the specific implementation steps are as follows.

[0067] 1. Complete missing trajectory

[0068] The LSTM algorithm can learn the historical features of time series data and predict future trends. This model has been widely used in trajectory prediction, and this invention introduces the LSTM algorithm to predict trajectories in occluded areas.

[0069] Before completing the missing trajectories in the occluded areas, the incomplete trajectories are first divided into a set of trajectories before occlusion. and occlusion trajectory set

[0070] 1) Predict the trajectory of the occluded area

[0071] The trajectory of each vehicle in the obstructed area is predicted. For example, the trajectory of the Kth vehicle in front of the obstructed area is... in Will The LSTM prediction network is trained using the input data, and the LSTM structure used for training the network is as follows: Figure 4 As shown.

[0072] The acceleration in the occlusion region is predicted using a cyclic iterative method, and the predicted acceleration sequence is used to calculate the trajectory sequence using equation (1). The vehicle's state at time t-1 is used as input to predict the state at time t. This iterative update process is performed until the trajectory sequence is obtained m seconds after the occlusion occurs. Here, t represents a time frame, and m is the ratio of the occlusion distance to the average speed of the vehicle before the occlusion occurs (in seconds).

[0073] The predicted complete trajectory set is

[0074]

[0075] 2) Match the predicted trajectory with the ID jump trajectory

[0076] To maintain consistency of trajectory IDs before and after occlusion, the predicted trajectory needs to be... Trajectory set after occlusion area To perform matching, here Let A' represent the trajectory sequence of the h-th car whose ID jumps after occlusion. This problem is transformed into a classic assignment problem with K tasks and N jobs, where each task must be assigned to one worker, all tasks must be performed, and the assignment objective is to minimize the total cost, where A' represents the task and B represents the worker. k，h Let represent that the k-th task is executed by the h-th worker, where k∈A′, h∈B, and represents a value of 1 if assigned and 0 otherwise. The main constraint of this assignment problem is... This means that each task can only be performed by one worker. This means that each worker can only perform the intended task; It refers to the range of values ​​a variable can take; C k，h It is the cost when employee h performs task k; x k，h It is the decision value for matching the k-th trajectory in A′ with the h-th trajectory in B; x is determined by x k，h The decision vector is composed of these components. The objective function of the classic assignment problem is to find the minimum cost while satisfying the above constraints, as shown in equation (2).

[0077]

[0078] In this invention, Euclidean distance is used to construct the matching cost C between the target trajectory and the trajectory to be matched. k，h . and Used to calculate matching cost C k，h ; It is the average value of the global time frame corresponding to the last second of the predicted trajectory A′ in the occluded area and the average value of the trajectory sequence; It is the average value of the time frame and trajectory sequence corresponding to the first second in trajectory set B. The cost constraint function established by Euclidean distance is shown in equation (3).

[0079]

[0080] The cost constraint function represents the correlation between two trajectories in both time and space; the smaller the adjustment cost, the greater the correlation, and vice versa. The Hungarian algorithm was used to solve this problem.

[0081] 3) Trajectory splicing

[0082] The predicted trajectory will be determined based on the matching scheme obtained from the Hungarian algorithm. Trajectory after occlusion area Perform splicing. Then, determine the integrity of the trajectory using the following rule: if A′ i If the last frame of B equals the first frame of B, then the trajectory is complete; if A′ i If the last frame of A' is smaller than the first frame of B, then linear interpolation is used to interpolate the missing frames; if A' i If the last frame of B is greater than the first frame of B, then delete A'. i The trajectory corresponding to the extra frames in the middle.

[0083] 2. Noise Trajectory Filtering Algorithm

[0084] To address the severe noise problem in CV output trajectories, a SFPF noise reduction algorithm is proposed, which integrates an interactive behavior model in microscopic traffic flow and a particle filter algorithm. This algorithm introduces a particle filter algorithm with an unrestricted noise distribution assumption as the trajectory fusion unit, a social force model as the state update equation in the particle filter, and the CV output trajectory as the input value to the observation update equation. The algorithm flow is as follows: Figure 5 As shown, it contains three nested loops, and the specific functions and implementations of each loop are as follows:

[0085] In the outer loop, to avoid systematic errors caused by noise from other vehicle trajectories to the reconstructed vehicle, multiple rounds of noise reduction processing are required for all noisy trajectories in the system. After each loop, the corresponding ID trajectory in the original noisy trajectory set is updated using the new trajectory set, and the iteratively stable result is selected as the final output. This method recommends more than 200 loops.

[0086] The middle loop performs noise reduction processing on all vehicles, and the noise reduction result will replace the corresponding vehicle ID trajectory in the environmental traffic flow.

[0087] The inner loop fuses the CV trajectory and the social force-generated trajectory frame by frame. At the end of each loop, the state equation module updates the vehicle motion state according to the social force model, and the measurement method module reads the original noise trajectory state values. This algorithm uses a particle filter to fuse the CV trajectory and the social force-generated trajectory. This process not only preserves the shape characteristics of the CV trajectory but also satisfies driving behavior constraints. Furthermore, the outer loop uses iterative convergence and multiple iterations to find the optimal value, effectively avoiding random errors. The parameters involved in the social force model can be calibrated using a small amount of real trajectory data; a genetic algorithm can be selected for the calibration.

[0088] This method involves two existing theoretical models: one is the particle filter algorithm; the other is the social force model, which are explained in detail below.

[0089] 1) Particle Filter Algorithm

[0090] The state equations and measurement equations of the particle filter system are as follows:

[0091] x k =f k (x k-1 v k -1) (4)

[0092] z k =h k (x k u k (5)

[0093] Where: x k The system's state value at time k; z k f is the observation at time k; k It is the state transition function of the system state; v k-1 and u k These represent the variances of the state transition noise and observation noise distributions, respectively. A value of v is recommended. k-1 =0.5, u k =1.0.

[0094] In this invention, a social force model is used to replace f. k h k This represents the system's measurement function, which is replaced here by the CV output.

[0095] The state probability density distribution can be approximated by the empirical probability distribution shown in equation (6).

[0096]

[0097] Where δ(·) is a Dirac function, f(x) 0：h The expected value is shown in equation (7):

[0098] E(f(x 0：k ))=∫f(x 0：k )p(x 0：k |z 1：k )dx 0：k (7)

[0099] The integral operation can be approximated by the following form of estimation:

[0100]

[0101] The Monte Carlo method transforms the problem of calculating integrals into a probability transition accumulation process with a finite number of sample points, implemented using an importance sampling algorithm. If the probability distribution equals p(x... k |z 1：k ) and probability density distribution q(x) 0：k |z 1：k If is a known, easily sampled distribution function, then equation (7) can be transformed into equation (10):

[0102]

[0103]

[0104] Among them, w k (x 0：k () is a non-normalized weighted average of importance; These are non-standardized importance weights; It is q(x) 0：k |z 1：k An independent sampling point in ).

[0105]

[0106] Importance sampling is a commonly used Monte Carlo integration method, where p(x) k |z 1：k Before any observation can be estimated, all observations must be available at time k. Each time a new observation is received, the importance weights of the entire state sequence must be recalculated.

[0107] 2) State update based on social force model

[0108] The inner loop of SFPF constructs a state update equation driven by an environmental perception and social force model to update the vehicle's motion state, ensuring that the trajectory conforms to driving behavior constraints. The motion state of frame t is updated based on the surrounding environment state of frame t-1, including the vehicle state "car" in the environment at time t. i，t ={x it y it ;v x，it vy，it a x，it a y，it}, Non-motorized vehicle status Bike k，t ={x kt y kt v x，kt v y，kt a x，kt a y，kt Ebike j，t ={x jt y jt v x，jt v y，jt a x，jt a y，jt} and road environmental constraints Border l，t ={l o , l ro The target vehicle perceives its environmental state, determines the interaction object, and uses the social force model to calculate the driving force, interaction force, and boundary constraint force to form a resultant force. The target vehicle's motion state is updated frame by frame. The acceleration in the next frame is determined by the force applied to the reconstructed target in the previous frame, which in turn affects the motion position in the next frame. Equation (12) is the social force model used. The force acting on the non-motorized vehicle at time t consists of 5 parts, where ξ represents the random term. The model parameter values ​​need to be calibrated using 5 minutes of real trajectory data.

[0109]

[0110] (a) Self-driving force

[0111] During vehicle operation, the goal is to reach the destination at the optimal speed and along the shortest path. When the vehicle deviates from the optimal speed, it automatically adjusts the magnitude and direction of acceleration under the influence of driving force to approach the ideal speed. The calculation process is shown in equation (13):

[0112]

[0113] In the formula, v is the expected speed of vehicle b at time t (m / s); b (t) The current speed of vehicle b at time t (m / s); τ b The average time required to change the relaxation time from the current speed to the periodic speed; a b Represents the expected acceleration; m b It is the mass of vehicle b.

[0114] (b) Interaction forces between non-motorized vehicles

[0115] Non-motorized vehicle sports are affected by other non-motorized vehicles in the line of sight. On the one hand, non-motorized vehicles are subject to repulsive forces, changing their speed to avoid collisions; on the other hand, they are also affected by overtaking forces, changing their direction to avoid collisions. This can be expressed as follows:

[0116]

[0117] In the formula, It is the repulsive force between adjacent non-motorized vehicles j at time t, in order to avoid a collision; The overtaking force is used to characterize the flexible behavior of non-motorized vehicle b when encountering obstacles or congestion.

[0118] Repulsive force:

[0119] The repulsive force between non-motorized vehicles is similar to the psychological repulsive force between pedestrians to meet the need for safe space, reflecting the cyclist's awareness of avoiding collisions with others. The repulsive force of non-motorized vehicle j on non-motorized vehicle b is shown in equation (15).

[0120]

[0121] In the formula: The minor semi-axis of the ellipse drawn with the location of non-motorized vehicle b as its focus; (v b -v j ) is the speed of non-motorized vehicle b relative to j; U3 and R3 represent the scaling factor and the distance sensitivity coefficient, respectively; r bj It is the vector from non-motorized vehicle j to b; Δt is the time step; n bj It is the direction vector from non-motorized vehicle j to b.

[0122] Overtaking ability:

[0123] Non-motorized vehicles are flexible in their movement and tend to change direction to overtake obstacles rather than remain stationary when they encounter them. This behavior is characterized by the overtaking force shown in Equation (16).

[0124]

[0125] In the formula, k is the proportionality coefficient; It is perpendicular to And with (v) b -v j The unit vectors with smaller included angles. The forces between non-motorized vehicles are similar to those in the pedestrian social force model, both depending on the perceived perspective. A person behind or to the side of a non-motorized vehicle has a much smaller influence on the vehicle than a person in front of it. Therefore, the same direction-dependent weight ω is introduced. bj .

[0126]

[0127] In the formula: λ b The range of variation is 0 << λ b <<1; It is the angle between the position vector of non-motorized vehicle b relative to j and the direction of travel of non-motorized vehicle b. The modified expression is:

[0128] (c) Boundary forces

[0129] When the distance between the non-motorized vehicle and the lane boundary is less than a certain value, the boundary force acts on the non-motorized vehicle to prevent it from exceeding the boundary. Its magnitude is related to the distance between the non-motorized vehicle and the road, as shown in equation (18):

[0130]

[0131] In the formula, U4 is the proportionality coefficient of the boundary repulsion force; r bw R4 is the distance between the non-motorized vehicle position and the lane boundary; R4 is the sensitivity coefficient of the boundary repulsion force to distance; n bw It refers to the unit vector pointing from the center of the non-motorized vehicle to the position on the lane closest to the non-motorized vehicle.

[0132] (d) Interaction forces with motor vehicles

[0133] When a motor vehicle enters the field of vision laterally, the non-motorized vehicle will experience a lateral force. The magnitude of the force depends on the lateral distance between the non-motorized and motorized vehicles and the speed of the motorized vehicle. The lateral force is calculated as shown in equation (19):

[0134]

[0135] In the formula, U6 and R6 are the parameters to be estimated; r bc It is the distance between the motor vehicle and the target vehicle; ||v c || refers to the speed of the motor vehicle, n bc This represents the unit vector of the interaction forces between motor vehicles and non-motor vehicles.

[0136] If there is a motor vehicle in front of the non-motorized vehicle, the non-motorized vehicle will experience a force that slows it down:

[0137]

[0138] s = s0 + v b Δt (21)

[0139] In the formula, b c s0 is the comfortable deceleration speed for non-motorized vehicles; v is the safe distance for non-motorized vehicles; bΔt is the target non-motorized vehicle speed; d is the reaction time; bc This refers to the distance between motor vehicles and non-motor vehicles, n v A unit vector representing the deceleration caused by the motor vehicle.

[0140] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for reconstructing the trajectory of non-motorized vehicles in a shared space between motorized and non-motorized vehicles, characterized in that, The method steps include: The initial trajectory is extracted using the CV algorithm; LSTM prediction is used to complete the trajectory missing due to occlusion in the initial trajectory, and the Hungarian algorithm is used to match the trajectories before and after occlusion to complete the complete trajectory. A three-layer cyclic social force particle filter algorithm, which incorporates a social force model representing non-motorized vehicle driving behavior into a particle filter framework, is used to denoise the completed trajectory and obtain the reconstructed trajectory. The described cyclic social force particle filter algorithm integrates the interactive behavior model in microscopic traffic flow with the particle filter algorithm. It uses the particle filter algorithm with the unrestricted noise distribution assumption as the trajectory fusion unit, the social force model as the state update equation in the particle filter, and the CV trajectory as the input value of the observation update equation. It contains three layers of loops, and the specific functions and implementations of each layer are as follows: The outer loop performs multiple rounds of noise reduction processing on all noise trajectories. The new round will execute the noise reduction algorithm based on the results of the previous round. After the multiple rounds of loops are completed, the most stable result of the iteration is selected as the final result. The middle loop performs noise reduction processing on all vehicles and replaces the corresponding vehicle ID trajectory in the environmental traffic flow with the noise reduction results; The inner loop fuses the CV trajectory and the social force-generated trajectory frame by frame. At the end of one loop, the state equation module updates the vehicle motion state according to the social force model and reads the original noise trajectory state value.

2. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 1, characterized in that, The missing trajectory is completed by constructing a prediction model using a two-layer LSTM structure. The input of the prediction model is the trajectory acceleration sequence before occlusion, and the output of the prediction model is the missing trajectory caused by occlusion. The two layers of LSTM are used to predict longitudinal acceleration and lateral acceleration, respectively.

3. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 1, characterized in that, The goal of the aforementioned trajectory matching before and after occlusion is to obtain a complete trajectory. It considers the consistency characteristics of the trajectories before and after occlusion, and also considers fusing the temporal and spatial feature information of the trajectories before and after occlusion. The objective function for matching is: In the formula, It is the first in the predicted trajectory set of the occluded area. k The first trajectory and the occluded trajectory set h The decision value for matching; For matching costs; The matching cost The equation is constructed using the Euclidean distance method and solved using the Hungarian algorithm, as shown below. In the formula, It is the average value of the global time frame corresponding to the last second of the predicted trajectory set of the occluded area and the average value of the trajectory sequence; It is the average value of the time frame corresponding to the first second in the trajectory set after occlusion and the average value of the trajectory sequence.

4. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 1, characterized in that, Based on the matching scheme obtained from the Hungarian algorithm, the predicted trajectory set and the trajectory set after the occlusion area are concatenated, and the integrity of the trajectory is judged to obtain the complete trajectory; the rules for the integrity judgment are as follows: If the last frame of the predicted trajectory set is equal to the first frame of the trajectory set after the occlusion area, then it is a complete trajectory. If the last frame of the predicted trajectory set is smaller than the first frame of the trajectory set after the occlusion area, then the missing frame is interpolated using the linear interpolation method. If the last frame of the predicted trajectory set is larger than the first frame of the trajectory set after the occlusion area, then delete the trajectory corresponding to the extra frame in the predicted trajectory set.

5. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 1, characterized in that, The described cyclic social force particle filter algorithm uses a social force model to measure and update the CV trajectory in the fusion algorithm. The state update equation is shown below: In the formula, Indicates self-driving force; This represents the interaction forces between non-motorized vehicles; Indicates boundary forces; Indicates the force applied by the motor vehicle; The error is random; the parameters of the social force model are calibrated using a genetic algorithm.

6. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 5, characterized in that, The expression for the self-driving force is shown in the following formula: In the formula, It is a vehicle b exist t Expected speed at any given moment; vehicle b exist t Current velocity at any given moment; The average time required to change the relaxation time from the current speed to the cycle speed; Represents expected acceleration; It is a vehicle b The quality.

7. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 5, characterized in that, The expression for the interaction force between non-motorized vehicles is shown in the following formula: In the formula, Adjacent non-motorized vehicles j In time t The repulsive force, For overtaking power, For the same directional dependency weights; Among them, non-motorized vehicles j For non-motorized vehicles b The expression for the repulsive force is shown in the following equation: In the formula, Therefore, non-motorized vehicles b The minor semi-axis of the ellipse drawn at the focal point; Non-motorized vehicles b relatively j speed; and These represent the scaling factor and the sensitivity coefficient to distance, respectively. From non-motorized vehicles j arrive b ; It is the time step; From non-motorized vehicles j arrive b The direction vector; The overtaking force is shown in the following formula: In the formula, k It is the proportionality coefficient; It is perpendicular to And with Unit vectors with smaller included angles; The same direction-dependent weights The expression is as follows: In the formula, The range of variation is ; Non-motorized vehicles b Compared to j Position vector and non-motorized vehicles b The angle between the directions of travel.

8. The method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 5, characterized in that, The magnitude of the boundary force is related to the distance between the non-motorized vehicle and the lane boundary, as shown in the following formula: In the formula, It is the proportionality coefficient of the magnitude of the boundary repulsion force; It is the distance between the position of the non-motorized vehicle and the lane boundary; It is the sensitivity coefficient of boundary repulsion force to distance; It refers to the unit vector pointing from the center of the non-motorized vehicle to the position on the lane closest to the non-motorized vehicle.

9. A method for reconstructing the trajectory of non-motorized vehicles in a shared space according to claim 5, characterized in that, The expression for the force acting on the motor vehicle is shown in the following formula: In the formula, and The parameter to be estimated; It is the distance between the motor vehicle and the target vehicle; This refers to the speed of a motor vehicle. A unit vector representing the interaction force between motor vehicles and non-motor vehicles; If there is a motor vehicle in front of the non-motorized vehicle, the non-motorized vehicle will experience a force that slows it down: In the formula, This is the comfortable deceleration speed for non-motorized vehicles; This refers to the safe distance for non-motorized vehicles. It is the target non-motorized vehicle speed; It is the reaction time; This refers to the distance between motor vehicles and non-motor vehicles. It is the unit vector of deceleration affected by motor vehicles.

Images