Real-time collision detection method based on driving intention and trajectory prediction

Abstract

The invention discloses a real-time collision detection method based on driving intention and trajectory prediction, and the method comprises the steps: constructing an intention-based trajectory prediction frame which does not need deep learning, constructing a driving intention prediction model based on features extracted from a non-intersection scene and an intersection scene, and carrying out the prediction of the driving intention based on a tree-shaped Parzen Estimation (Tree-structured Parzen Estimation). The method comprises the following steps: completing training of a driving intention prediction model by using a Bayesian optimization algorithm based on TPE, predicting driving intentions of obstacle vehicles around a controlled vehicle by using the driving intention prediction model obtained by training, and predicting driving tracks of the obstacle vehicles around the controlled vehicle for a non-intersection scene and an intersection scene respectively. Whether the controlled vehicle and the obstacle vehicle will collide or not is judged through comparison and calculation of the driving tracks of the controlled vehicle and the obstacle vehicle, and the calculation real-time performance of collision detection is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of prediction planning for unmanned vehicles, and in particular relates to a real-time collision detection method based on driving intention and trajectory prediction. Background Art

[0002] Collision detection is an important task in autonomous driving. Accurate and reliable collision detection is essential to ensure the safety of autonomous vehicles. At present, many deep learning-based algorithms have achieved good performance in the field of trajectory prediction, but some of them are not suitable for deployment in engineering projects. The reasons are: on the one hand, considering the many factors that affect the future vehicle trajectory, including surrounding vehicles, pedestrians, lane markings and complex multimodal features, the forward reasoning time of the model often cannot meet the real-time requirements due to the excessive number of parameters involved. In autonomous driving, even the slightest delay can lead to catastrophic accidents. The unmet demand for real-time performance limits the applicability of the most advanced methods. On the other hand, most mainstream trajectory prediction algorithms can only predict the trajectory of the next 3-5 seconds, and this short-term prediction trajectory is not enough to meet the safety requirements of autonomous vehicles in some high-speed scenarios. In addition, due to the inevitable deviation of the predicted trajectory of surrounding vehicles, especially when the prediction field of view is long, simply relying on determining whether two vehicles will collide at the same time for collision detection may lead to unreliable collision detection results. Summary of the invention

[0003] In view of this, the present invention provides a real-time collision detection method based on driving intention and trajectory prediction, which realizes the real-time detection of the collision risk between a controlled vehicle and an obstacle vehicle.

[0004] The present invention provides a real-time collision detection method based on driving intention and trajectory prediction, which specifically includes the following steps:

[0005] Step 1: Obtain vehicle trajectory information and the true value of the vehicle driving intention annotated therein to construct a training sample data set;

[0006] Step 2: A driving intention prediction model is established based on a single-layer LSTM network of a non-deep network to predict the driving intention of obstacle vehicles around the controlled vehicle. The input of the driving intention prediction model is shown in the following formula:

[0007]

[0008] Among them, I i is the historical trajectory sequence feature of the i-th obstacle vehicle, is the feature of the i-th obstacle vehicle at time t; the feature of the non-intersection scene is expressed as As shown below:

[0009]

[0010] in, is the distance from the starting point of the road to the current vehicle position along the road reference line at time tk, is the offset perpendicular to the current position along the road reference line at time tk, is the offset of the obstacle vehicle relative to the current heading angle of the controlled vehicle at time tk, is the speed of the obstacle vehicle at time tk, is the acceleration value of the obstacle vehicle at time tk;

[0011] The feature representation of the intersection scene is As shown below:

[0012]

[0013] in, and is the position coordinate of the ith obstacle vehicle in the plane rectangular coordinate system at time tk in the given UTM coordinate system, is the speed of the obstacle vehicle along the X axis at time tk, is the speed of the obstacle vehicle along the Y axis at time tk;

[0014] Step 3: Using the Bayesian optimization algorithm based on tree Parzen estimation to obtain the optimal hyperparameters of the driving intention prediction model, and completing the training of the driving intention prediction model;

[0015] Step 4: In actual use, the vehicle trajectory information of the obstructing vehicle is used as the input of the trained driving intention prediction model, and the driving intention of the obstructing vehicle is calculated by the driving intention prediction model. Then, the driving trajectory of the obstructing vehicle is predicted for non-intersection scenes and intersection scenes respectively, and the trajectory prediction range is the set time period;

[0016] Step 5: Determine the directional bounding box of the controlled vehicle and the obstacle vehicle associated with it, and expand the length and width of the directional bounding box by k times; then introduce the collision time difference to represent the time interval for the controlled vehicle and the obstacle vehicle to reach the same position. If the collision time difference between the two is less than the set threshold δ, it is determined that the two vehicles will collide, otherwise it is determined that the two vehicles will not collide.

[0017] Furthermore, the prediction method for the driving trajectory of vehicles with obstacles in non-intersection scenarios is as follows:

[0018] For a vehicle traveling in a straight line, it is assumed that it maintains uniform acceleration along the road direction and the speed does not exceed the maximum speed threshold, and it continues to approach the center line of the lane in the direction perpendicular to the road. The process is expressed as shown in the following formula:

[0019]

[0020] Among them, s t is the distance traveled by the obstacle vehicle along the road reference line at time t, s t+1 is the distance traveled by the obstacle vehicle along the road reference line at time t+1, v s is the projection of the obstacle vehicle’s velocity in the S direction, a s is the projection of the obstacle vehicle’s acceleration in the S direction, Δt is the prediction time interval, is the L coordinate of the projection point of the obstacle vehicle on the center line at time t, k is the speed of the obstacle vehicle approaching the center line, l t is the offset of the obstacle vehicle along the road reference line perpendicular to the current position at time t, l t+1 The offset of the obstacle vehicle along the road reference line perpendicular to the current position at time t+1;

[0021] For vehicles with lane-changing intentions, the fourth-order polynomial s(t) and the fifth-order polynomial l(t) are used to model the trajectories in the S and L directions, as shown in the following equations:

[0022]

[0023] s(t) is the distance traveled by the obstacle vehicle along the road reference line, l(t) is the distance traveled by the obstacle vehicle in the direction perpendicular to the road reference line; s(0) = s0 means that the state of s0 is substituted to construct the first expression for solving s(t), Indicates substitution The state of constructing the second expression for solving s(t), and so on, solving the coefficients of the s(t) polynomial; l(0) = l0 means substituting the state of l0 to construct the first expression, Indicates substitution The state of the obstacle vehicle is used to construct the second expression, and the coefficients of the l(t) polynomial are solved in the same way; s0, l0 and their higher-order derivatives are calculated from the current state of the obstacle vehicle, T1 is the lane change time, and the candidate lane change trajectory of the obstacle vehicle is calculated by sampling T1 in the Frenet coordinate system; then the cost function shown in the following formula is used to select the optimal trajectory:

[0024] cost lc (T1) = w1T1 + w2max{a n}

[0025] Among them, w1 and w2 are weight parameters, max{a n} is the maximum normal acceleration in the trajectory, and the trajectory with the smallest cost function value is selected as the lane change trajectory.

[0026] Furthermore, the method for predicting the driving trajectory of the obstacle vehicle in the intersection scene is as follows:

[0027] The turning or straight line trajectory in the X and Y directions in the UTM coordinate system is modeled as shown below:

[0028]

[0029] Among them, x(0)=x0 means substituting the state of x0 to construct the first expression for solving x(t), Indicates substitution The state of l0 is used to construct the second expression for x(t), and so on, the coefficients of the polynomial x(t) can be solved; y(0) = y0 means substituting the state of l0 to construct the first expression, Indicates substitution The state of the second expression is constructed, and the coefficients of the y(t) polynomial are solved in the same way; T1 is the time required for the obstacle vehicle to pass the intersection, x exit and exit is the UTM coordinate of the first point on the center line of the exit lane; a series of candidate trajectories are obtained by sampling T1, and the trajectory with the smallest cost function value is selected as the optimal trajectory.

[0030] Furthermore, the dataset NGSIM is used to train the driving intention prediction model for non-intersection scenarios, and the dataset INTERACTION is used to train the driving intention prediction model for intersection scenarios.

[0031] Furthermore, the driving intention includes turning left, turning right or going straight.

[0032] Furthermore, the set time period is 8 seconds.

[0033] Furthermore, the vehicle trajectory information includes vehicle position, speed, acceleration and heading angle.

[0034] Beneficial effects:

[0035] 1. The present invention constructs an intention-based trajectory prediction framework that does not require deep learning, constructs a driving intention prediction model based on features extracted from non-intersection scenarios and intersection scenarios, and uses a Bayesian optimization algorithm based on a tree-structured Parzen Estimator (TPE) to complete the training of the driving intention prediction model. The trained driving intention prediction model is used to predict the driving intention of obstacle vehicles around the controlled vehicle, and then the driving trajectories of obstacle vehicles around the controlled vehicle are predicted for non-intersection scenarios and intersection scenarios respectively. By comparing and calculating the driving trajectories of the controlled vehicle and the obstacle vehicle, it is determined whether the controlled vehicle and the obstacle vehicle will collide, thereby effectively improving the computational real-time performance of collision detection.

[0036] 2. The present invention introduces time redundancy and space redundancy in the collision detection process, which effectively avoids the problem of overly stringent computational requirements of the current trajectory-based collision detection algorithm. It can effectively cope with the computational challenges that may be faced in real scenarios and improve the method's adaptability to the requirements of autonomous driving collision detection projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the network structure of a driving intention prediction model constructed in a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention.

[0038] Figure 2 A schematic diagram of driving intention in a non-intersection scenario determined by a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention.

[0039] Figure 3 A schematic diagram of trajectory prediction output in a non-intersection scenario determined by a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention.

[0040] Figure 4 A schematic diagram of driving intention in an intersection scenario determined by a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention.

[0041] Figure 5 A schematic diagram of trajectory prediction output in an intersection scenario determined by a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention.

[0042] Figure 6 A schematic diagram of an intersection collision scenario after introducing spatial and temporal redundancy into a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention.

[0043] Figure 7 The figure is a schematic diagram of a collision warning output obtained when a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention is used without adding redundancy.

[0044] Figure 8 A schematic diagram of adding redundant collision warning output to a real-time collision detection method based on driving intention and trajectory prediction provided by the present invention. DETAILED DESCRIPTION

[0045] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0046] The present invention provides a real-time collision detection method based on driving intention and trajectory prediction. The basic idea is to build a driving intention prediction model based on features extracted from non-intersection scenes and intersection scenes respectively, and use a Bayesian optimization algorithm based on tree Parzen estimation to complete the training of the driving intention prediction model. The trained driving intention prediction model is used to predict the driving intention of obstacle vehicles around a controlled vehicle. Then, the driving trajectories of obstacle vehicles around the controlled vehicle are predicted for non-intersection scenes and intersection scenes respectively. By comparing and calculating the driving trajectories of the controlled vehicle and the obstacle vehicle, it is determined whether the controlled vehicle and the obstacle vehicle will collide.

[0047] The present invention provides a real-time collision detection method based on driving intention and trajectory prediction, which specifically includes the following steps:

[0048] Step 1: Obtain vehicle trajectory information and the true value of the vehicle driving intention annotated therein to construct a training sample dataset.

[0049] Specifically, the present invention can use NGSIM and INTERACTION in the public data set as training sample data sets, wherein the data set NGSIM can be used to train the driving intention prediction model for non-intersection scenarios, and the data set INTERACTION can be used to train the driving intention prediction model for intersection scenarios. The present invention limits the number of samples of each class in the data set to the minimum among the number of samples of all classes to alleviate data imbalance, and divides the training set, test set, and validation set into 7:1.5:1.5.

[0050] Step 2: Establish a driving intention prediction model based on a single-layer LSTM network with a non-deep network. The model structure is as follows: Figure 1 As shown, the driving intention of the obstacle vehicles located around the controlled vehicle is predicted, including turning left, turning right, or going straight.

[0051] The input of the driving intention prediction model constructed by the present invention is shown in the following formula:

[0052]

[0053] Among them, I i is the historical trajectory sequence feature of the i-th obstacle vehicle, is the feature of the i-th obstacle vehicle at time t. The feature of the non-intersection scene is expressed as As shown below:

[0054]

[0055] in, is the distance from the starting point of the road to the current vehicle position along the road reference line at time tk, is the offset perpendicular to the current position along the road reference line at time tk, is the offset of the obstacle vehicle relative to the current heading angle of the controlled vehicle at time tk, is the speed of the obstacle vehicle at time tk, is the acceleration value of the obstacle vehicle at time tk.

[0056] The feature representation of the intersection scene is As shown below:

[0057]

[0058] in, and is the position coordinate of the ith obstacle vehicle in the plane rectangular coordinate system at time tk in the given UTM coordinate system, is the speed of the obstacle vehicle along the X axis at time tk, is the speed of the obstacle vehicle along the Y axis at time tk.

[0059] The output of the driving intention prediction model is the probability of occurrence of various intentions corresponding to different scenarios, and the intention with the highest probability among all feasible intentions is taken as the final intention. The predicted driving intention in non-intersection scenarios is as follows: Figure 2 As shown in the figure, a positive integer 1 indicates that the vehicle changes lanes to the left but does not enter the target lane, and a positive integer 2 indicates that the vehicle changes lanes to the left and has entered the target lane; a negative integer -1 indicates that the vehicle changes lanes to the left but does not enter the target lane, and a negative integer -2 indicates that the vehicle changes lanes to the left and has entered the target lane. The predicted driving intention in the intersection scenario is as follows Figure 4 As shown, a positive integer 1 indicates that the vehicle intends to change lanes to the left at the intersection, and a positive integer 2 indicates that the vehicle intends to turn left at the intersection; a negative integer -1 indicates that the vehicle intends to change lanes to the left at the intersection, and a negative integer -2 indicates that the vehicle intends to turn left at the intersection.

[0060] Step 3: Use the Bayesian optimization algorithm based on the Tree-structured Parzen Estimator (TPE) to obtain the optimal hyperparameters of the driving intention prediction model and complete the training of the driving intention prediction model.

[0061] For example, the optimal hyperparameters of the driving intention prediction model obtained by the TPE-based Bayesian optimization algorithm are shown in Table 1:

[0062] Table 1 Optimal hyperparameters for driving intention prediction model

[0063] TPE-LSTM Hyperparameters Non-intersection scenarios Intersection scene <![CDATA[Fully connected layer hidden layer size hs1]]> 30 45 <![CDATA[Hidden layer size hs2 of the LSTM layer]]> 45 45 Number of LSTM layers ln 1 3 Learning rate α 0.001 0.0001 Number of training rounds ep 260 180 Batch size bs 4096 2048

[0064] Step 4: In actual use, the vehicle trajectory information of the obstructing vehicle is used as the input of the trained driving intention prediction model. The driving intention prediction model calculates the driving intention of the obstructing vehicle, and then predicts the driving trajectory of the obstructing vehicle for non-intersection scenes and intersection scenes respectively. The trajectory prediction range is a set period, which can be 8 seconds. Among them, the vehicle trajectory information includes vehicle position, speed, acceleration and heading angle.

[0065] Specifically, in non-intersection scenarios, the present invention generates trajectories in the S direction (along the road direction) and the L direction (perpendicular to the road direction) of the Frenet coordinate system, including two situations: vehicles traveling in a line and vehicles with the intention of changing lanes, as follows:

[0066] For a vehicle traveling in a straight line, it is assumed that it maintains uniform acceleration in the S direction and the speed does not exceed the maximum speed threshold, and it continues to approach the center line of the lane in the L direction. The process is expressed as shown in the following formula:

[0067]

[0068] Among them, s t is the distance traveled by the obstacle vehicle along the road reference line at time t, s t+1 is the distance traveled by the obstacle vehicle along the road reference line at time t+1, v s is the projection of the obstacle vehicle’s velocity in the S direction, a s is the projection of the obstacle vehicle’s acceleration in the S direction, Δt is the prediction time interval, is the L coordinate of the projection point of the obstacle vehicle on the center line at time t, k is the speed of the obstacle vehicle approaching the center line, l t is the offset of the obstacle vehicle along the road reference line perpendicular to the current position at time t, l t+1 It is the offset of the obstacle vehicle along the road reference line perpendicular to the current position at time t+1.

[0069] For vehicles with lane-changing intentions, the fourth-order polynomial s(t) and the fifth-order polynomial l(t) are used to model the trajectories in the S direction and the L direction, respectively, as shown in the following equations:

[0070]

[0071] This set of equations describes the process of solving the two state variables s(t) and l(t) and their first-order and second-order derivatives. s(t) is the distance traveled by the obstacle vehicle along the road reference line, and l(t) is the distance traveled by the obstacle vehicle in the direction perpendicular to the road reference line. s(0) = s0 means that the state of s0 is substituted to construct the first expression for solving s(t). Indicates substitution The state of s(t) is used to construct the second expression for solving s(t), and so on, the coefficients of the polynomial s(t) can be solved. Similarly, l(0) = l0 means that the state of l0 is substituted to construct the first expression, Indicates substitution The state of s0 is used to construct the second expression, and the coefficients of the l(t) polynomial can be solved by analogy. s0, l0 and their higher-order derivatives can all be calculated based on the current state of the obstacle vehicle. T1 is the lane change time. By sampling T1, the candidate lane change trajectory of the obstacle vehicle can be calculated in the Frenet coordinate system. For the above multiple candidate lane change trajectories, the cost function shown in the following formula is used to select the optimal trajectory:

[0072] cost lc (T1) = w1T1 + w2max{a n} (6)

[0073] Among them, w1 and w2 are weight parameters, max{a n} is the maximum normal acceleration in the trajectory, and the trajectory with the smallest cost function value is selected as the lane change trajectory.

[0074] In the intersection scenario, the third polynomials x(t) and y(t) are used to model the turning or straight line trajectories in the X and Y directions in the UTM coordinate system as shown below:

[0075]

[0076] Among them, x(0)=x0 means substituting the state of x0 to construct the first expression for solving x(t), Indicates substitution The state of l0 is used to construct the second expression for solving x(t), and so on, the coefficients of the polynomial x(t) can be solved. Similarly, y(0) = y0 means that the state of l0 is substituted to construct the first expression, Indicates substitution The state of the second expression is constructed, and the coefficients of the y(t) polynomial can be solved by analogy. T1 is the time required for the obstacle vehicle to pass the intersection, x exit and exit is the UTM coordinate of the first point on the center line of the exit lane. A series of candidate trajectories are obtained by sampling T1, and the optimal trajectory in the intersection scenario is calculated by formula (6) in the same way. The longest trajectory prediction time in the present invention is 8 seconds. When the prediction time is less than 8 seconds, the trajectory is extended to 8 seconds by using formula (4) to extend the straight line trajectory.

[0077] The driving trajectory predicted by the present invention in the non-intersection scenario is as follows: Figure 3 As shown in Figure 2, the driving trajectory in the intersection scenario is as follows: Figure 5 shown.

[0078] Step 5: Determine the oriented bounding boxes (OBB) of the controlled vehicle and the obstacle vehicle associated with it, and expand the length and width of the determined oriented bounding box by k times to set the collision space redundancy; after determining the oriented bounding boxes of the controlled vehicle and the obstacle vehicle, introduce the time difference of collision (TDC) to represent the time interval for the controlled vehicle and the obstacle vehicle to reach the same position. If the collision time difference between the two is less than the set threshold δ, it is determined that the two vehicles will collide, otherwise it is determined that the two vehicles will not collide.

[0079] Existing collision detection algorithms usually only evaluate whether the OBBs of two trajectories intersect at the same time. However, due to the inevitable errors in path and speed prediction, the reliability of the algorithm is affected. In order to solve this problem, the present invention enhances the reliability of traditional collision detection algorithms by introducing spatial and temporal redundancy. Specifically, spatial redundancy is achieved by expanding the length and width of the OBB by k times, which can improve the detection accuracy to a certain extent; temporal redundancy is achieved by introducing the collision time difference TDC, that is, if the collision time difference value is less than the threshold, it is considered that the two vehicles will collide on the future space-time trajectory.

[0080] Intuitively, the accuracy of the predicted trajectory will deviate in position and velocity over time. Therefore, when designing spatial and temporal redundancy, the present invention sets the OBB amplification factor k and the TDC threshold δ t , and the values ​​of these two variables will increase with time, so the present invention designs a linear saturation function r(t) to represent redundancy, as shown in the following formula:

[0081] r(t)=max{at,r max} (8)

[0082] Among them, a is the linear coefficient, r max is the maximum value of redundancy, max{at,r max} takes values ​​of at and r max The larger value of Figure 6 As shown in the figure, after introducing redundant judgment, taking the intersection scene as an example, the vehicle can stably output the collision point and perceive it in advance, which greatly improves the problem of insufficient collision risk perception in other collision algorithms. Figure 7 As shown in , 0 means there is no collision in the current frame, and 1 means there is a collision in the current frame. After adding the redundant setting, Figure 8As shown, 0 indicates no collision in the current frame, and 1 indicates a collision in the current frame. It can be observed that the controlled vehicle has a more stable perception of the collision with the surrounding obstacle vehicles, which can avoid the problem of frequent collision warning mutations at the collision critical point.

[0083] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time collision detection method based on driving intention and trajectory prediction, characterized in that: The specific steps include: Step 1: Obtain vehicle trajectory information and the true value of the vehicle driving intention annotated therein to construct a training sample data set; Step 2: A driving intention prediction model is established based on a single-layer LSTM network of a non-deep network to predict the driving intention of obstacle vehicles around the controlled vehicle. The input of the driving intention prediction model is shown in the following formula: Among them, I i is the historical trajectory sequence feature of the i-th obstacle vehicle, is the feature of the i-th obstacle vehicle at time t; the feature of the non-intersection scene is expressed as As shown below: in, is the distance from the starting point of the road to the current vehicle position along the road reference line at time tk, is the offset perpendicular to the current position along the road reference line at time tk, is the offset of the obstacle vehicle relative to the current heading angle of the controlled vehicle at time tk, is the speed of the obstacle vehicle at time tk, is the acceleration value of the obstacle vehicle at time tk; The feature representation of the intersection scene is As shown below: in, and is the position coordinate of the ith obstacle vehicle in the plane rectangular coordinate system at time tk in the given UTM coordinate system, is the speed of the obstacle vehicle along the X axis at time tk, is the speed of the obstacle vehicle along the Y axis at time tk; Step 3: Using the Bayesian optimization algorithm based on tree Parzen estimation to obtain the optimal hyperparameters of the driving intention prediction model, and completing the training of the driving intention prediction model; Step 4: In actual use, the vehicle trajectory information of the obstructing vehicle is used as the input of the trained driving intention prediction model, and the driving intention of the obstructing vehicle is calculated by the driving intention prediction model. Then, the driving trajectory of the obstructing vehicle is predicted for non-intersection scenes and intersection scenes respectively, and the trajectory prediction range is the set time period; Step 5: Determine the directional bounding box of the controlled vehicle and the obstacle vehicle associated with it, and expand the length and width of the directional bounding box by k times; then introduce the collision time difference to represent the time interval for the controlled vehicle and the obstacle vehicle to reach the same position. If the collision time difference between the two is less than the set threshold δ, it is determined that the two vehicles will collide, otherwise it is determined that the two vehicles will not collide.

2. The real-time collision detection method according to claim 1, characterized in that: The prediction method for the driving trajectory of vehicles with obstacles in non-intersection scenarios is as follows: For a vehicle traveling in a straight line, it is assumed that it maintains uniform acceleration along the road direction and the speed does not exceed the maximum speed threshold, and it continues to approach the center line of the lane in the direction perpendicular to the road. The process is expressed as shown in the following formula: Among them, s t is the distance traveled by the obstacle vehicle along the road reference line at time t, s t+1 is the distance traveled by the obstacle vehicle along the road reference line at time t+1, v s is the projection of the obstacle vehicle’s velocity in the S direction, a s is the projection of the obstacle vehicle’s acceleration in the S direction, Δt is the prediction time interval, is the L coordinate of the projection point of the obstacle vehicle on the center line at time t, k is the speed of the obstacle vehicle approaching the center line, l t is the offset of the obstacle vehicle along the road reference line perpendicular to the current position at time t, l t+1 The offset of the obstacle vehicle along the road reference line perpendicular to the current position at time t+1; For vehicles with lane-changing intentions, the fourth-order polynomial s(t) and the fifth-order polynomial l(t) are used to model the trajectories in the S and L directions, as shown in the following equations: s(t) is the distance traveled by the obstacle vehicle along the road reference line, l(t) is the distance traveled by the obstacle vehicle in the direction perpendicular to the road reference line; s(0) = s0 means that the state of s0 is substituted to construct the first expression for solving s(t), Indicates substitution The state of constructing the second expression for solving s(t), and so on, solving the coefficients of the s(t) polynomial; l(0) = l0 means substituting the state of l0 to construct the first expression, Indicates substitution The state of the obstacle vehicle is used to construct the second expression, and the coefficients of the l(t) polynomial are solved in the same way; s0, l0 and their higher-order derivatives are calculated from the current state of the obstacle vehicle, T1 is the lane change time, and the candidate lane change trajectory of the obstacle vehicle is calculated by sampling T1 in the Frenet coordinate system; then the cost function shown in the following formula is used to select the optimal trajectory: cost lc (T1)=w1T1+w2max{a n } Among them, w1 and w2 are weight parameters, max{a n } is the maximum normal acceleration in the trajectory, and the trajectory with the smallest cost function value is selected as the lane change trajectory.

3. The real-time collision detection method according to claim 1, characterized in that: The method for predicting the driving trajectory of an obstacle vehicle in an intersection scenario is as follows: The turning or straight line trajectory in the X and Y directions in the UTM coordinate system is modeled as shown below: Among them, x(0)=x0 means substituting the state of x0 to construct the first expression for solving x(t), Indicates substitution The state of l0 is used to construct the second expression for x(t), and so on, the coefficients of the polynomial x(t) can be solved; y(0) = y0 means substituting the state of l0 to construct the first expression, Indicates substitution The state of the second expression is constructed, and the coefficients of the y(t) polynomial are solved in the same way; T1 is the time required for the obstacle vehicle to pass the intersection, x exit and exit is the UTM coordinate of the first point on the center line of the exit lane; a series of candidate trajectories are obtained by sampling T1, and the trajectory with the smallest cost function value is selected as the optimal trajectory.

4. The real-time collision detection method according to claim 1, characterized in that: The NGSIM dataset is used to train the driving intention prediction model for non-intersection scenarios, and the INTERACTION dataset is used to train the driving intention prediction model for intersection scenarios.

5. The real-time collision detection method according to claim 1, characterized in that: The driving intention includes turning left, turning right or going straight.

6. The real-time collision detection method according to claim 1, characterized in that: The set time period is 8 seconds.

7. The real-time collision detection method according to claim 1, characterized in that: The vehicle trajectory information includes vehicle position, speed, acceleration and heading angle.

Images