AEB braking methods, devices, equipment, and media for pedestrian occlusion scenarios

Abstract

This application relates to an AEB (Autonomous Emergency Braking) method, device, equipment, and medium for pedestrian occlusion scenarios. The method includes: modeling the problem of protecting occluded pedestrians in vehicle active safety as a partially observable Markov decision process; solving for the optimal state-action utility function corresponding to each state combination in the state space; outputting an optimal policy set; calculating the vehicle's braking time based on the current vehicle state; calculating the average collision time based on the optimal policy set and the Cartesian product between the future states of the vehicle and the future states of the pedestrian in the state space; assessing the collision risk based on the average collision time and the vehicle's braking time; if the collision risk exceeds a threshold, the AEB system triggers emergency braking; otherwise, gradient braking is performed according to the action corresponding to the pedestrian state with the maximum belief value in the optimal policy set. This method can improve the triggering effect of the AEB system on occluded pedestrians, reduce the probability of false triggering and missed triggering, and improve the braking safety and comfort of the AEB system.

Description

Technical Field

[0001] This application relates to the field of vehicle active safety technology, and in particular to an AEB braking method, device, equipment and medium for pedestrian occlusion scenarios. Background Technology

[0002] With the rapid development of autonomous driving technology, the Automatic Emergency Braking (AEB) system has emerged as a key component. Utilizing advanced sensors such as millimeter-wave radar, lidar, and cameras, the AEB system can comprehensively perceive the position, speed, and trajectory of surrounding vehicles, pedestrians, and non-motorized vehicles. Upon detecting a potential collision, it initiates braking in advance, effectively reducing the risk of collision and greatly improving road safety, especially playing a crucial role in protecting pedestrians and non-motorized vehicles.

[0003] Traditional AEB systems primarily rely on a single metric such as Time-of-Collision (TTC) to trigger braking. This method determines the braking timing by predicting the pedestrian's trajectory and comparing it with a preset threshold. However, due to limitations in sensor accuracy and environmental uncertainties, this method is often overly conservative, easily leading to frequent false triggers of the AEB system and affecting driving comfort. Furthermore, while some data-driven AEB methods improve adaptability through multi-source data learning, they generally rely on instantaneous state information, resulting in significant TTC calculation errors and a substantial risk of false or missed triggers. This is especially true in scenarios where the pedestrian is obscured, as the pedestrian's state cannot be directly obtained, drastically reducing the reliability of these methods. Summary of the Invention

[0004] Therefore, it is necessary to provide an AEB braking method, device, equipment, and medium for pedestrian occlusion scenarios to address the aforementioned technical problems.

[0005] An AEB braking method for pedestrian occlusion scenarios, the method comprising:

[0006] The problem of protecting obscured pedestrians in active vehicle safety is modeled as a partially observable Markov decision process. This partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes discretized vehicle and pedestrian states. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes a longitudinal gradient braking strategy for the vehicle. The observation space is used to distinguish between unobscured and obscured areas using sensor observations. The transition model represents the state transition probabilities for both the vehicle and the pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. The discount factor is used to balance immediate and future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated using a belief representation, and the belief is iteratively updated using a Bayesian formula.

[0007] Solve for the optimal state-action utility function corresponding to each state combination in the state space, and output the optimal policy set; the optimal policy set includes the action with the highest utility value corresponding to the state combination; the state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief;

[0008] The vehicle braking time is calculated based on the current vehicle state. The average collision time is calculated based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrians in the state space. The collision risk is assessed based on the average collision time and the vehicle braking time.

[0009] If the collision risk exceeds the threshold, the AEB system triggers emergency braking; otherwise, gradient braking is performed according to the pedestrian state corresponding to the maximum value of the belief set in the optimal strategy.

[0010] In one embodiment, the method further includes: modifying the transfer model based on the braking execution results of the AEB system.

[0011] In one embodiment, the comfort constraints include imposing a high penalty for colliding with a pedestrian and a low penalty for excessive deceleration; the safety constraints include rewarding the maintenance of a safe distance and constant speed.

[0012] In one embodiment, the observation model includes: pedestrian state detection values ​​in unobstructed areas follow a normal distribution around the true state, while pedestrians in obstructed areas are undetectable.

[0013] In one embodiment, the expression for the Bayesian formula iterative update of beliefs is:

[0014]

[0015] Where b′(s′) is the updated belief, O(o|s,a) is the observation model, T(s′|a,s) is the transition model, b(s) is the current belief, s is the current pedestrian state, a is the vehicle state, and s′ is the pedestrian state after the transition.

[0016] In one embodiment, the vehicle braking time is calculated using the vehicle's initial speed and its maximum deceleration capability.

[0017] In one embodiment, calculating the average collision time based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrian in the state space includes: obtaining multiple state combinations based on the Cartesian product between the future states of each vehicle and the future states of the pedestrian in the state space; calculating the estimated proportion of overlap between the future states of the vehicle and the future states of the pedestrian based on the actions, vehicle states, pedestrian states, and transition models in the optimal strategy set corresponding to the state combinations, and obtaining the collision probability; and calculating the average collision time based on the collision time of the state combinations whose collision probability is greater than the collision threshold.

[0018] An AEB braking device for pedestrian occlusion scenarios, the device comprising:

[0019] The problem construction module is used to model the problem of protecting occluded pedestrians in vehicle active safety as a partially observable Markov decision process. This partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes discretized vehicle and pedestrian states. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes a longitudinal gradient braking strategy for the vehicle. The observation space is used to distinguish between unoccluded and occluded areas using sensor observations. The transition model represents the state transition probabilities for both the vehicle and the pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. The discount factor is used to weigh immediate and future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated using a belief representation, and the belief is iteratively updated using a Bayesian formula.

[0020] The strategy optimization module is used to solve for the optimal state-action utility function corresponding to each state combination in the state space and output the optimal strategy set; the optimal strategy set includes the action with the highest utility value corresponding to the state combination; the state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief;

[0021] The risk assessment module is used to calculate the vehicle braking time based on the current vehicle state, calculate the average collision time based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of pedestrians in the state space, and assess the collision risk based on the average collision time and the vehicle braking time.

[0022] The AEB braking module is used to trigger emergency braking if the collision risk is greater than a threshold; otherwise, it performs gradient braking according to the pedestrian state corresponding to the maximum value of the belief in the optimal strategy set.

[0023] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0024] The problem of protecting obscured pedestrians in active vehicle safety is modeled as a partially observable Markov decision process. This partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes discretized vehicle and pedestrian states. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes a longitudinal gradient braking strategy for the vehicle. The observation space is used to distinguish between unobscured and obscured areas using sensor observations. The transition model represents the state transition probabilities for both the vehicle and the pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. The discount factor is used to balance immediate and future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated using a belief representation, and the belief is iteratively updated using a Bayesian formula.

[0025] Solve for the optimal state-action utility function corresponding to each state combination in the state space, and output the optimal policy set; the optimal policy set includes the action with the highest utility value corresponding to the state combination; the state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief;

[0026] The vehicle braking time is calculated based on the current vehicle state. The average collision time is calculated based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrians in the state space. The collision risk is assessed based on the average collision time and the vehicle braking time.

[0027] If the collision risk exceeds the threshold, the AEB system triggers emergency braking; otherwise, gradient braking is performed according to the pedestrian state corresponding to the maximum value of the belief set in the optimal strategy.

[0028] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0029] The problem of protecting obscured pedestrians in active vehicle safety is modeled as a partially observable Markov decision process. This partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes discretized vehicle and pedestrian states. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes a longitudinal gradient braking strategy for the vehicle. The observation space is used to distinguish between unobscured and obscured areas using sensor observations. The transition model represents the state transition probabilities for both the vehicle and the pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. The discount factor is used to balance immediate and future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated using a belief representation, and the belief is iteratively updated using a Bayesian formula.

[0030] Solve for the optimal state-action utility function corresponding to each state combination in the state space, and output the optimal policy set; the optimal policy set includes the action with the highest utility value corresponding to the state combination; the state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief;

[0031] The vehicle braking time is calculated based on the current vehicle state. The average collision time is calculated based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrians in the state space. The collision risk is assessed based on the average collision time and the vehicle braking time.

[0032] If the collision risk exceeds the threshold, the AEB system triggers emergency braking; otherwise, gradient braking is performed according to the pedestrian state corresponding to the maximum value of the belief set in the optimal strategy.

[0033] The aforementioned AEB braking method, device, equipment, and medium for pedestrian occlusion scenarios model the protection problem of occluded pedestrians as a partially observable Markov decision process. It utilizes state-space discretization to process key parameters of both the vehicle and the pedestrian, and iteratively updates beliefs using Bayesian formulas to quantify the probability distribution of pedestrian states, accurately capturing uncertainties in occlusion scenarios. By solving for the optimal state-action utility function to generate a strategy set, and combining vehicle braking time and mean collision time to assess risk, it achieves flexible control at the decision-making level. It coordinates gradient braking strategies with AEB emergency braking, forming a full-scenario coverage from smooth driving to emergency avoidance through gradient actions guided by the maximum belief value and emergency braking as a fallback. This complementary function of POMDP and AEB ensures comfort through graded braking while maintaining a safety baseline through emergency braking, ultimately achieving a braking effect that is both comfortable and safe. The embodiments of this invention can improve the triggering effect of the AEB system on occluded pedestrians, reduce the probability of false triggering and missed triggering, and improve the braking safety and comfort of the AEB system. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating an AEB braking method for a pedestrian occlusion scenario in one embodiment.

[0035] Figure 2 This is a structural block diagram of an AEB braking device for a pedestrian occlusion scenario in one embodiment;

[0036] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] In one embodiment, such as Figure 1 As shown, an AEB braking method for pedestrian occlusion scenarios is provided, including the following steps:

[0039] Step 102: Model the problem of protecting obscured pedestrians in vehicle active safety as a partially observable Markov decision process.

[0040] Partially observable Markov decision processes include state space, action space, observation space, transition model, observation model, reward function, and discount factor.

[0041] The state space includes discretized vehicle and pedestrian states. Vehicle states include longitudinal velocity, and pedestrian states include relative longitudinal distance, lateral offset, speed, and direction. The action space includes the vehicle's longitudinal gradient braking strategy. The observation space uses sensor observations to distinguish between unobstructed and occluded areas. The transition model represents the state transition probabilities for both the vehicle and pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. A discount factor is used to balance immediate and future rewards to calculate the utility value. The probability distribution of pedestrian states is estimated using beliefs, which are iteratively updated using Bayes' theorem.

[0042] Partially observable Markov decision processes (POMDPs) model pedestrian occlusion scenarios as partially observable environments. They use belief states to represent probabilistic estimates of the pedestrian's true state, making them suitable for scenarios where sensors cannot fully observe the pedestrian's state. By discretizing the pedestrian's motion parameters, the continuous state space is transformed into a finite set of states, facilitating subsequent processing. The relative longitudinal position refers to the relative distance between the pedestrian and the vehicle in the longitudinal direction of travel, with the vehicle's current position as the reference point, and forward (the pedestrian being in front of the vehicle) as the positive direction, ranging from 0 to 50 meters. Lateral offset refers to the distance the pedestrian deviates from the center line of the vehicle's trajectory in the direction perpendicular to the vehicle's travel direction, with the center line of the vehicle's trajectory as the reference line, negative to the left and positive to the right, ranging from -5 meters to 5 meters. Direction refers to the angle between the pedestrian's movement direction and the vehicle's travel direction, with the vehicle's travel direction as the 0-degree reference point, negative angles for deflection to the left of the vehicle's travel direction and positive angles for deflection to the right, ranging from -90 degrees to 90 degrees (covering scenarios such as pedestrians crossing the road laterally and moving diagonally). The vehicle's longitudinal gradient braking strategy includes constant speed (0 m / s). 2 ) and -1m / s 2 up to -3m / s 2 The deceleration (at intervals of 0.5 m / s²) 2 This system provides graded braking capabilities through a longitudinal gradient braking strategy, avoiding the limitations of traditional AEB systems that rely solely on emergency braking. It also improves the accuracy of estimating occluded pedestrian states by dynamically fusing new observations with historical states based on beliefs.

[0043] Step 104: Solve for the optimal state-action utility function corresponding to each state combination in the state space, and output the optimal policy set.

[0044] The optimal policy set comprises the actions with the highest utility value corresponding to the state combinations. State combinations are obtained by combining the vehicle's state with the pedestrian's state predicted by beliefs. The long-term cumulative reward of taking a specific action in each state is quantified using a state-action utility function to guide optimal decision-making. The optimal policy set forms a decision table covering all possible scenarios, supporting real-time response.

[0045] Step 106: Calculate the vehicle braking time based on the current vehicle state, calculate the average collision time based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrian in the state space, and assess the collision risk based on the average collision time and the vehicle braking time.

[0046] Self-braking time is the minimum time required to assess emergency braking of a self-vehicle. By considering the mean time to impact (MTI) across all possible collision scenarios and their probabilities, a more comprehensive risk assessment is provided.

[0047] Step 108: If the collision risk is greater than the threshold, the AEB system triggers emergency braking; otherwise, gradient braking is performed according to the pedestrian state corresponding to the maximum value of the optimal strategy concentration belief.

[0048] Braking based on risk level can reduce unnecessary emergency braking while retaining emergency braking capability in high-risk situations, ensuring a safety baseline.

[0049] In the aforementioned AEB braking method for pedestrian occlusion scenarios, the problem of protecting occluded pedestrians is modeled as a partially observable Markov decision process. Key parameters of both the vehicle and the pedestrian are processed using state space discretization, and beliefs are iteratively updated using Bayesian formulas to quantify the probability distribution of pedestrian states, accurately capturing uncertainties in occlusion scenarios. By solving for the optimal state-action utility function to generate a strategy set, and combining vehicle braking time and mean collision time to assess risk, flexible control at the decision-making level is achieved. Gradient braking strategies are coordinated with AEB emergency braking, forming a full-scenario coverage from smooth driving to emergency avoidance through gradient actions guided by the maximum belief value and emergency braking as a fallback. This complementary function of POMDP and AEB ensures comfort through graded braking while maintaining a safety baseline through emergency braking, ultimately achieving a braking effect that is both comfortable and safe. This embodiment of the invention can improve the triggering effect of the AEB system on occluded pedestrians, reduce the probability of false triggering and missed triggering, and improve the braking safety and comfort of the AEB system.

[0050] In one embodiment, the method further includes: correcting the transition model based on the braking execution results of the AEB system. In this embodiment, the AEB system feeds back data from the actual braking process, such as the actual deceleration effect and collision avoidance situation, to the POMDP planner for dynamic optimization. Specifically, by comparing the braking execution results with the predicted values ​​of the transition model, the state transition probabilities of pedestrians and vehicles are corrected, making the model more consistent with the actual scenario. The transition model includes vehicle state transitions and pedestrian state transitions. Vehicle state transitions can be estimated using motion models such as the CV (constant speed) model, CS (constant steering) model, or uniform acceleration model. Pedestrian state transitions can be estimated using pedestrian motion models such as uniformly selecting any acceleration or the CV model.

[0051] In one embodiment, comfort constraints include imposing a high penalty for colliding with a pedestrian and a low penalty for excessive deceleration; safety constraints include rewarding maintaining a safe distance and traveling at a constant speed.

[0052] In one embodiment, the observation model includes: pedestrian state detection values ​​in unobstructed areas follow a normal distribution around the true state, while pedestrians in obstructed areas are undetectable.

[0053] In one embodiment, the expression for iteratively updating beliefs using the Bayesian formula is:

[0054]

[0055] Where b′(s′) is the updated belief, O(o|s,a) is the observation model, T(s′|a,s) is the transition model, b(s) is the current belief, s is the current pedestrian state, a is the vehicle state, and s′ is the pedestrian state after the transition.

[0056] In one embodiment, the vehicle braking time is calculated using the vehicle's initial speed and its maximum deceleration capability.

[0057] In one embodiment, calculating the average collision time based on the optimal policy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrian in the state space includes: obtaining multiple state combinations based on the Cartesian product between the future states of each vehicle and the future states of the pedestrian in the state space; calculating the estimated proportion of overlap between the future states of the vehicle and the future states of the pedestrian based on the actions, vehicle states, pedestrian states, and transition models in the optimal policy set corresponding to the state combinations, and obtaining the collision probability; and calculating the average collision time based on the collision time of state combinations with collision probabilities greater than the collision threshold.

[0058] In one specific embodiment, the specific steps for implementing the method of the present invention include:

[0059] Step 1: Model the problem of protecting occluded pedestrians using active vehicle safety methods as a partially observable Markov decision process (POMDP). The overall framework can be represented by a 7-tuple (S, A, o, T, O, R, γ), which includes the state space S, action space A, observation space o, transition model T, observation model o, reward function R, and discount factor γ. Belief b is updated using a Bayesian update formula, based on a discrete Bayesian updater, combined with continuous observations to update the discretized belief. Specifically, the state space S contains vehicle and pedestrian information, such as position and speed; the action space A includes longitudinal acceleration curve control; the observation space o includes the vehicle's perception of the environment; the transition model T includes the state transitions of the vehicle and pedestrian; and the reward function R includes penalty collision and reward for maintaining speed.

[0060] Step 2: Design a POMDP planner and use the offline QMDP (Q-based Markov Decision Process) method to compute the optimal policy. This method is based on the assumption that the state is fully observable after a time step and solves for the optimal state-action utility function through a value iteration algorithm. The planner controls the vehicle's actions in the longitudinal direction, and the longitudinal control policy actions include constant speed driving and braking deceleration of varying intensities. All variables in the state space are discretized, generating a large number of state combinations.

[0061] Step 3: Integrate the POMDP planner into the Automatic Emergency Braking (AEB) system. The AEB system works in conjunction with the POMDP planner. The AEB system uses several state combinations generated by the POMDP planner to assess collision risk. If all actions within the POMDP planner's action space cannot avoid a collision, the AEB system will trigger emergency braking. The AEB system operates at a higher frequency to quickly detect emergencies. Information input to the AEB system includes pedestrian position and speed, as well as the vehicle's trajectory. The AEB system then calculates the braking time to braking (TTB) and the probability of collision. If the risk is too high, emergency braking is triggered, and the vehicle will perform the maximum deceleration allowed by its own capabilities.

[0062] Furthermore, step 1 is a Markov decision process with state uncertainty. The current observations of the pedestrian's state received by the vehicle are imperfect, so the modeling of the pedestrian's state is not entirely accurate. However, by utilizing subsequent observations of the pedestrian from the past and their subsequent actions, the vehicle can gradually approach a realistic state model. The vehicle's beliefs are represented by probability distributions over the basic states, and these beliefs are updated based on the vehicle's observations and actions. If the state space is discrete, or satisfies certain linear Gaussian assumptions, then accurate belief updates can be performed. If these assumptions are not met, approximations based on linearization or sampling can also be used.

[0063] When the vehicle is in state s∈S, if action a is taken, the state transitions to state s′ with probability T(s′|s,a)=Pr(s′|s,a). In POMDP, the vehicle has uncertain knowledge about the environmental state; therefore, the vehicle forms a belief based on its internal understanding of the current state. This belief b can be updated after performing action a and observing the current state o using the following formula:

[0064]

[0065] Where T(s′|a,s) denotes the state transition function, starting from the initial belief distribution before the vehicle takes any action or makes any observation, and then using a discrete Bayesian updater that updates the discretized beliefs through continuous observations of the measurements. The solution to POMDP is an optimal policy that maximizes the expected discounted sum of the immediate rewards for any given belief.

[0066] Furthermore, step 2 will solve the POMDP problem modeled in step 1. In POMDP, the goal is to select actions that maximize reward accumulation while interacting with the environment. However, the state of the POMDP problem is not directly observable, which requires the vehicle to use its past behavior and observation history information to update its beliefs.

[0067] Step 2.1: Determine the vehicle's motion space. The longitudinal motion space of the vehicle, such as constant speed driving and braking of different magnitudes, is represented by a set of different deceleration magnitudes: {0 m / s 2 -1m / s 2 -1.5m / s 2 -2m / s 2 -2.5m / s 2 -3m / s 2}

[0068] Step 2.2: Define the state space, which contains all possible variables for solving the problem. It includes information about the vehicle and pedestrians, represented in the Frenet coordinate system. The vehicle's longitudinal speed ranges from 0 km / h to 40 km / h. The pedestrian's longitudinal range relative to the vehicle is less than 50 meters, and their lateral range is less than ±5 meters. The pedestrian's speed is less than 6 km / h, and their direction varies between ±90 degrees. The state space contains 48 possible vehicle speeds and 20 possible longitudinal positions, 10 possible lateral positions, 4 possible speeds, and 9 possible directions for the pedestrian. By multiplying all possible vehicle and pedestrian states, the total number of states is 3.4 * 10^35. 5 .

[0069] Step 2.3: Determine the transfer model. The transfer model of the vehicle depends on the current action and state of the vehicle and is determined by a certain target tracking motion model, such as the CA (constant acceleration) model. For the transfer model of the pedestrian, the further position of the pedestrian can be calculated based on the assumption that the pedestrian can uniformly choose any acceleration. The upper limit of the pedestrian's speed is 2m / s.

[0070] Step 2.4: Determine the observation model. The observation model describes the vehicle's perception of the state space. Assuming that the vehicle's position and velocity can be completely observed, the observation model can be described as follows:

[0071] 1. Objects in unobstructed areas will always be detected.

[0072] 2. Objects obstructing behind obstacles will not be detected.

[0073] If a pedestrian is detected, the pedestrian's position, speed, and direction are all normally distributed around its true state.

[0074] Step 2.5: Determine the reward model. The reward model defines the goal of the POMDP planner. The vehicle will be penalized for hitting a pedestrian. At the same time, to avoid excessive intervention, longitudinal actions (deceleration) will also be penalized. By selecting different penalty and reward values, the behavior of the POMDP planner can be adjusted to balance collision avoidance and vehicle driving efficiency (average speed).

[0075] Furthermore, in step 3, the AEB system works in conjunction with the POMDP planner. The AEB system uses information provided by the POMDP planner to determine whether the vehicle is at risk of collision. If a collision cannot be avoided, the AEB system will trigger emergency braking, that is, send an emergency braking signal to the vehicle's braking system to cause the vehicle to brake to its maximum extent.

[0076] Step 3.1: Receive information from the POMDP planner and calculate the required braking time for the vehicle based on that information:

[0077] Here, start_velocity is the initial speed of the vehicle, and max_deceleration is the maximum deceleration capability of the vehicle.

[0078] Step 3.2: Based on the state space information provided by the POMDP planner, the collision probability can be calculated by calculating the Cartesian product of each state set and determining the estimated proportion of overlap between the pedestrian state and the future vehicle state. If the probability is high, the average collision time is calculated based on the combination of states that will cause a collision. If this time is less than the braking time calculated in Step 3.1, it means that the maximum deceleration in the vehicle's action space in Step 2.1 cannot meet the safety requirements, and an emergency stop command will be issued.

[0079] It is understood that the POMDP planner included in the AEB system of this invention has excellent robustness to uncertainties in pedestrian states, such as considering hidden pedestrians behind occluded pedestrians. The POMDP planner is designed to perform comfortable deceleration within a given deceleration range and is responsible for considering uncertainties caused by occlusion. When one side of the road is obstructed, the POMDP planner adjusts the vehicle speed. Integrating the POMDP planner into the AEB system enhances the POMDP strategy, allowing the AEB system, which utilizes the full braking force of the vehicle, to operate more effectively. When an unavoidable collision occurs, the AEB system is responsible for implementing forceful intervention, thus the entire AEB system offers multiple deceleration options to cover a wider range of scenarios.

[0080] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0081] In one embodiment, such as Figure 2 As shown, an AEB braking device for pedestrian occlusion scenarios is provided, comprising:

[0082] Problem construction module 202 is used to model the problem of protecting occluded pedestrians in vehicle active safety as a partially observable Markov decision process. The partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes the discretized vehicle state and pedestrian state. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes the vehicle's longitudinal gradient braking strategy. The observation space is used to distinguish between unoccluded and occluded areas using sensor observations. The transition model represents the state transition probabilities of the vehicle and pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort constraints and safety constraints. The discount factor is used to weigh immediate rewards against future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated through belief representation. The belief is iteratively updated using Bayes' theorem.

[0083] The strategy optimization module 204 is used to solve the optimal state-action utility function corresponding to each state combination in the state space and output the optimal policy set. The optimal policy set includes the action with the highest utility value corresponding to the state combination. The state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief.

[0084] Risk assessment module 206 is used to calculate the vehicle braking time based on the current vehicle state, calculate the average collision time based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of pedestrians in the state space, and assess the collision risk based on the average collision time and the vehicle braking time.

[0085] AEB braking module 208 is used to trigger emergency braking if the collision risk is greater than a threshold; otherwise, it performs gradient braking according to the pedestrian state corresponding to the maximum value of the optimal strategy concentration belief.

[0086] In one embodiment, it is also used to modify the transfer model based on the braking execution results of the AEB system.

[0087] In one embodiment, comfort constraints include imposing a high penalty for colliding with a pedestrian and a low penalty for excessive deceleration; safety constraints include rewarding maintaining a safe distance and traveling at a constant speed.

[0088] In one embodiment, the observation model includes: pedestrian state detection values ​​in unobstructed areas follow a normal distribution around the true state, while pedestrians in obstructed areas are undetectable.

[0089] In one embodiment, the expression for iteratively updating beliefs using the Bayesian formula is:

[0090]

[0091] Where b′(s′) is the updated belief, O(o|s,a) is the observation model, T(s′|a,s) is the transition model, b(s) is the current belief, s is the current pedestrian state, a is the vehicle state, and s′ is the pedestrian state after the transition.

[0092] In one embodiment, the vehicle braking time is calculated using the vehicle's initial speed and its maximum deceleration capability.

[0093] In one embodiment, the system is further configured to obtain multiple state combinations based on the Cartesian product between the future states of the vehicle and the future states of the pedestrian in the state space; calculate the estimated proportion of overlap between the future states of the vehicle and the future states of the pedestrian based on the actions, vehicle states, pedestrian states and transition models in the optimal policy set corresponding to the state combinations, and obtain the collision probability; and calculate the average collision time based on the collision time of the state combinations with collision probabilities greater than the collision threshold.

[0094] Specific limitations regarding AEB braking devices for pedestrian-occluded scenarios can be found in the above-mentioned limitations on AEB braking methods for pedestrian-occluded scenarios, and will not be repeated here. Each module in the aforementioned AEB braking device for pedestrian-occluded scenarios can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0095] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an AEB (Automatic Emergency Braking) method for pedestrian occlusion scenarios. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0096] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0097] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0098] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0099] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An AEB braking method for pedestrian occlusion scenarios, characterized in that, The method includes: The problem of protecting obscured pedestrians in active vehicle safety is modeled as a partially observable Markov decision process. This partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes discretized vehicle and pedestrian states. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes a longitudinal gradient braking strategy for the vehicle. The observation space is used to distinguish between unobscured and obscured areas using sensor observations. The transition model represents the state transition probabilities for both the vehicle and the pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. The discount factor is used to balance immediate and future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated using a belief representation, and the belief is iteratively updated using a Bayesian formula. Solve for the optimal state-action utility function corresponding to each state combination in the state space, and output the optimal policy set; the optimal policy set includes the action with the highest utility value corresponding to the state combination; the state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief; The vehicle braking time is calculated based on the current vehicle state. The average collision time is calculated based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of pedestrians in the state space. The collision risk is assessed based on the average collision time and the vehicle braking time. If the collision risk exceeds the threshold, the AEB system triggers emergency braking; otherwise, gradient braking is performed according to the pedestrian state corresponding to the maximum value of the belief set in the optimal strategy.

2. The method according to claim 1, characterized in that, The method further includes: The transfer model is modified based on the braking execution results of the AEB system.

3. The method according to claim 1, characterized in that, The comfort constraints include a high penalty for colliding with a pedestrian and a low penalty for excessive deceleration; the safety constraints include a reward for maintaining a safe distance and driving at a constant speed.

4. The method according to claim 1, characterized in that, The observation model includes: The pedestrian state detection values ​​in the unobstructed area follow a normal distribution around the true state, while pedestrians in the obstructed area cannot be detected.

5. The method according to claim 1, characterized in that, The expression for iteratively updating beliefs using the Bayesian formula is as follows: Where b′(s′) is the updated belief, O(o|s,a) is the observation model, T(s′|a,s) is the transition model, b(s) is the current belief, s is the current pedestrian state, a is the vehicle state, and s′ is the pedestrian state after the transition.

6. The method according to claim 1, characterized in that, The braking time of the vehicle is calculated using the initial speed of the vehicle and the maximum deceleration capability of the vehicle.

7. The method according to claim 1, characterized in that, The average collision time is calculated based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of the pedestrians in the state space, including: Multiple state combinations can be obtained by the Cartesian product between the future states of the vehicles and the future states of the pedestrians in the state space; Based on the actions, vehicle state, pedestrian state, and transition model in the optimal policy set corresponding to the state combination, calculate the estimated proportion of overlap between the future state of the vehicle and the future state of the pedestrian to obtain the collision probability. The average collision time is calculated based on the collision time of state combinations with a collision probability greater than the collision threshold.

8. An AEB braking device for pedestrian occlusion scenarios, characterized in that, The device includes: The problem construction module is used to model the problem of protecting occluded pedestrians in vehicle active safety as a partially observable Markov decision process. This partially observable Markov decision process includes a state space, action space, observation space, transition model, observation model, reward function, and discount factor. The state space includes discretized vehicle and pedestrian states. The vehicle state includes the vehicle's longitudinal velocity, and the pedestrian state includes the pedestrian's relative longitudinal distance, lateral offset, speed, and direction. The action space includes a longitudinal gradient braking strategy for the vehicle. The observation space is used to distinguish between unoccluded and occluded areas using sensor observations. The transition model represents the state transition probabilities for both the vehicle and the pedestrian. The observation model represents the probabilistic relationship between sensor observations and the actual state. The reward function includes comfort and safety constraints. The discount factor is used to weigh immediate and future rewards to calculate the utility value. The probability distribution of the pedestrian state is estimated using a belief representation, and the belief is iteratively updated using a Bayesian formula. The strategy optimization module is used to solve for the optimal state-action utility function corresponding to each state combination in the state space and output the optimal strategy set; the optimal strategy set includes the action with the highest utility value corresponding to the state combination; the state combination is obtained by combining the vehicle state and the pedestrian state predicted by belief; The risk assessment module is used to calculate the vehicle braking time based on the current vehicle state, calculate the average collision time based on the optimal strategy set and the Cartesian product between the future states of each vehicle and the future states of pedestrians in the state space, and assess the collision risk based on the average collision time and the vehicle braking time. The AEB braking module is used to trigger emergency braking if the collision risk is greater than a threshold; otherwise, it performs gradient braking according to the pedestrian state corresponding to the maximum value of the belief in the optimal strategy set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images