A v2i-based intersection vehicle behavior prediction system and method
By combining the V2I system with roadside and vehicle-to-everything (V2X) sensing devices and utilizing V2X communication and edge computing, the problems of unstable accuracy and high computational load in autonomous vehicle behavior prediction have been solved, achieving efficient vehicle behavior prediction even in the absence of historical trajectory data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XINTONG INST INNOVATION CENT FOR INTERNET OF VEHICLES (CHENGDU) CO LTD
- Filing Date
- 2022-12-29
- Publication Date
- 2026-06-09
AI Technical Summary
Existing autonomous vehicle behavior prediction technologies face significant challenges in predicting behavior when historical trajectory data of the target is lacking. The choice of algorithm model also affects the instability of accuracy, and the computational load is large, requiring high hardware resources.
A V2I-based intersection vehicle behavior prediction system is adopted. By combining roadside perception devices and vehicle perception devices, V2X communication is used to share driving intentions. Deep learning and multi-sensor fusion are combined with edge computing devices to reduce dependence on historical motion trajectories, improve prediction accuracy and reduce computation.
Improve prediction accuracy, reduce computational load and hardware resource requirements, and achieve more stable vehicle behavior prediction without target historical motion trajectory data.
Smart Images
Figure CN116052114B_ABST
Abstract
Description
Technical fields:
[0001] This invention belongs to the field of vehicle networking, specifically relating to a V2I-based system and method for predicting vehicle behavior at intersections. Background technology:
[0002] With the continuous advancement of science and technology, intelligent control technology, mainly based on computer technology and automation technology, has developed rapidly. Therefore, in recent years, research on autonomous intelligent vehicles has become one of the most concerned topics.
[0003] Behavior prediction is one of the core modules of autonomous driving systems. It is responsible for predicting the intentions and trajectories of perceived moving targets and judging the driving intentions of surrounding vehicles in order to ensure the driving safety of the autonomous driving system.
[0004] Existing behavior prediction technologies primarily rely on traditional vehicle-mounted sensing devices to receive historical motion trajectory information of targets. Based on multiple factors such as the target's historical motion trajectory and road environment constraints, deep learning algorithms are used to predict the target's behavior. However, this approach suffers from the following problems:
[0005] 1. Without historical data on the target's movement trajectory, prediction becomes very difficult.
[0006] 2. The selection of the prediction model directly affects the accuracy of the prediction, and indirectly leads to unstable accuracy.
[0007] 3. The single-vehicle prediction model has a large computational load and high requirements for the hardware resources of the prediction system. Summary of the Invention
[0008] To address the aforementioned issues, the primary objective of this invention is to provide a V2I-based vehicle behavior prediction system and method for intersections. This system utilizes V2X communication to share driving intentions, combining vehicle perception with roadside perception to improve prediction accuracy, reduce computational load and hardware resource requirements, and decrease reliance on the target's historical trajectory.
[0009] To achieve the above objectives, the technical solution of the present invention is as follows:
[0010] A V2I-based method for predicting vehicle behavior at intersections, the system comprising: a roadside sensing device (Sensor), an edge computing device (MEC), a roadside unit (RSU), and a vehicle vehicle (HV), wherein the roadside sensing device (Sensor), the edge computing device (MEC), the roadside unit (RSU), and the vehicle vehicle (HV) are connected in sequence.
[0011] Furthermore, the autonomous vehicle (HV) is equipped with vehicle sensing devices and a V2X communication unit (OBU), which communicates with the RSU in real time through the OBU. The vehicle sensing devices include cameras, radar, and GPS.
[0012] Furthermore, the roadside sensing device includes cameras and radar.
[0013] A V2I-based method for predicting vehicle behavior at intersections, comprising the following steps:
[0014] S1: The roadside sensing device will send the raw sensing data of the target vehicle (RV) and other surrounding targets to the MEC in real time;
[0015] S2: MEC uses a deep learning-based multi-sensor fusion perception method to perceive target information of RV and other surrounding targets based on the raw perception data in S1;
[0016] S3: MEC uses the RV's motion history and target information of other surrounding targets, combined with lane information in the high-precision map, and adopts a semantic map-based intersection prediction model to output the prior probability of each intention of the RV and predict the RV's driving intention.
[0017] S4: MEC samples the trajectory of each RV intent based on the intersection map data, sampling multiple possible motion trajectories. Then, it calculates the cost of each motion trajectory by considering the cost function of acceleration and centripetal acceleration. Based on the magnitude of the cost, it selects the most reasonable trajectory for each intent. Finally, it calculates a posterior probability by estimating the likelihood of the cost function and combining it with the prior probability, thus obtaining the final trajectory prediction probability.
[0018] S5: MEC will send RV target information, including intent prediction probability and motion prediction trajectory, to the autonomous vehicle HV via RSU;
[0019] S6: The autonomous vehicle (HV) matches the RV target information sensed by the roadside equipment sent by the MEC with the multi-target information sensed by the autonomous vehicle's sensing equipment, filters out its own behavior prediction for the RV, and performs weighted fusion of the roadside prediction results and vehicle-side prediction results based on RV distance and tracking time to calculate the final RV behavior prediction result, which is then used by the HV's decision planning module.
[0020] Furthermore, in step S2, the target information includes, but is not limited to, position, speed, direction of travel, and size.
[0021] Furthermore, in step S2, the specific implementation process of the deep learning multi-sensor fusion perception method is as follows: the target data perceived by multiple sensors is spatiotemporally aligned, the Hungarian algorithm is used to associate multiple targets, and the associated data is fused through Kalman filtering, ultimately achieving target-level fusion of multi-sensor perception data.
[0022] Furthermore, in step S3, the intersection prediction model based on semantic maps is an existing technology, derived from Apollo's obstacle behavior prediction technology.
[0023] Furthermore, the cost function, the method for selecting the cost magnitude, and the method for calculating the posterior probability in step S4 are all existing technologies, derived from Apollo's obstacle behavior prediction technology.
[0024] Furthermore, in step S6, the HV's vehicle perception devices include cameras, radar, and GPS.
[0025] Furthermore, in step S6, the specific method of weighted fusion is as follows: the driving intention of the vehicle RV at the intersection includes four types: going straight, turning left, turning right, and making a U-turn. The output of the behavior prediction is the probability of the four intentions and their corresponding driving trajectories.
[0026] The predicted straight-ahead probability P at the road end rs The probability of turning left, P rl The probability of turning right, P rr The probability of turning around, P ru , where P rs +P rl +P rr +P ru =1;
[0027] The vehicle-predicted probability P of going straight vs The probability of going straight, P vl The probability of turning right, P vr The probability of turning around, P vu , where P vs +P vl +P vr +P vu =1;
[0028] Let the weight of the road end be W. r The weight of the vehicle side is W. v W r +W v =1;
[0029] We perform weighted fusion of data from the roadside and vehicle-side, and calculate the fused probability:
[0030]
[0031] The weight W on the vehicle side v The calculation is related to the vehicle's sensing range and tracking time, specifically:
[0032] Based on the sensing range of the HV's autonomous vehicle sensing equipment and the RV's location, first determine whether the RV is within the effective sensing range of the autonomous vehicle sensing equipment:
[0033] If RV is not within the perception range, the weight is calculated using the following formula:
[0034] If the RV is within the sensing range, the maximum effective sensing distance of the vehicle sensing device is calculated as D based on the performance and installation location of the vehicle sensing device. s The distance D between RV and HV is calculated based on the position of RV. r The formula for calculating the weight is:
[0035] Compared with existing technologies, the advantages of this invention are: it enables driving intention sharing based on V2X communication, combines vehicle perception with roadside perception, improves prediction accuracy, and reduces computational load and hardware resource requirements; secondly, this prediction method can predict target trajectories even without historical motion trajectory data of the target, thus reducing dependence on the historical motion trajectory of the target. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the vehicle behavior prediction system for intersections according to the present invention.
[0037] Figure 2 This is a system workflow diagram of the present invention.
[0038] Figure 3 This is a diagram showing the sensing range of the HV vehicle sensing device of the present invention.
[0039] Figure 4 This is a diagram of the perception range of the front-view camera in this embodiment. Detailed Implementation
[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0041] Figures 1-4 As shown, the present invention is implemented as follows:
[0042] A V2I-based method for predicting vehicle behavior at intersections, the system comprising: a roadside sensing device (Sensor), an edge computing device (MEC), a roadside unit (RSU), and a vehicle vehicle (HV), wherein the roadside sensing device (Sensor), the edge computing device (MEC), the roadside unit (RSU), and the vehicle vehicle (HV) are connected in sequence.
[0043] Furthermore, the autonomous vehicle (HV) is equipped with vehicle sensing devices and a V2X communication unit (OBU), which communicates with the RSU in real time through the OBU. The vehicle sensing devices include cameras, radar, and GPS.
[0044] Furthermore, the roadside sensing device includes cameras and radar.
[0045] A V2I-based method for predicting vehicle behavior at intersections, comprising the following steps:
[0046] S1: The roadside sensing device will send the raw sensing data of the target vehicle (RV) and other surrounding targets to the MEC in real time;
[0047] S2: MEC uses a deep learning-based multi-sensor fusion perception method to perceive target information of RV and other surrounding targets based on the raw perception data in S1;
[0048] S3: MEC uses the RV's motion history and target information of other surrounding targets, combined with lane information in the high-precision map, and adopts a semantic map-based intersection prediction model to output the prior probability of each intention of the RV and predict the RV's driving intention.
[0049] S4: MEC samples the trajectory of each RV intent based on the intersection map data, sampling multiple possible motion trajectories. Then, it calculates the cost of each motion trajectory by considering the cost function of acceleration and centripetal acceleration. Based on the magnitude of the cost, it selects the most reasonable trajectory for each intent. Finally, it calculates a posterior probability by estimating the likelihood of the cost function and combining it with the prior probability, thus obtaining the final trajectory prediction probability.
[0050] S5: MEC will send RV target information, including intent prediction probability and motion prediction trajectory, to the autonomous vehicle HV via RSU;
[0051] S6: The autonomous vehicle (HV) matches the RV target information sensed by the roadside equipment sent by the MEC with the multi-target information sensed by the autonomous vehicle's sensing equipment, filters out its own behavior prediction for the RV, and performs weighted fusion of the roadside prediction results and vehicle-side prediction results based on RV distance and tracking time to calculate the final RV behavior prediction result, which is then used by the HV's decision planning module.
[0052] Furthermore, in step S2, the target information includes, but is not limited to, position, speed, direction of travel, and size.
[0053] Furthermore, in step S2, the specific implementation process of the deep learning multi-sensor fusion perception method is as follows: the target data perceived by multiple sensors is spatiotemporally aligned, the Hungarian algorithm is used to associate multiple targets, and the associated data is fused through Kalman filtering, ultimately achieving target-level fusion of multi-sensor perception data.
[0054] Furthermore, in step S3, the intersection prediction model based on semantic maps is an existing technology, derived from Apollo's obstacle behavior prediction technology.
[0055] Furthermore, the cost function, the method for selecting the cost magnitude, and the method for calculating the posterior probability in step S4 are all existing technologies, derived from Apollo's obstacle behavior prediction technology.
[0056] Furthermore, in step S6, the HV's vehicle perception devices include cameras, radar, and GPS.
[0057] Furthermore, in step S6, the specific method of weighted fusion is as follows: the driving intention of the vehicle RV at the intersection includes four types: going straight, turning left, turning right, and making a U-turn. The output of the behavior prediction is the probability of the four intentions and their corresponding driving trajectories.
[0058] The predicted straight-ahead probability P at the road end rs The probability of turning left, P rl The probability of turning right, P rr The probability of turning around, P ru , where P rs +P rl +P rr +P ru =1;
[0059] The vehicle-predicted probability P of going straight vs The probability of going straight, P vl The probability of turning right, P vr The probability of turning around, P vu , where P vs +P vl +P vr +P vu =1;
[0060] Let the weight of the road end be W. r The weight of the vehicle side is W. v W r +W v =1;
[0061] We perform weighted fusion of data from the roadside and vehicle-side, and calculate the fused probability:
[0062]
[0063] The weight W on the vehicle side v The calculation is related to the vehicle's sensing range and tracking time.
[0064] like Figure 3 As shown, each sensor has a limited sensing range and cannot detect targets outside that range. Furthermore, the accuracy of target detection decreases with increasing distance. The effective sensing range of vehicle-mounted sensors depends on the type, location, and number of sensors installed.
[0065] Taking the sensing range of a front-view camera as an example: Figure 4 As shown, based on the sensing range of the HV's autonomous vehicle sensing device and the RV's location, the first step is to determine whether the RV is within the effective sensing range of the autonomous vehicle sensing device:
[0066] If RV is not within the perception range, the weight is calculated using the following formula:
[0067] If the RV is within the sensing range, the maximum effective sensing distance of the vehicle sensing device is calculated as D based on the performance and installation location of the vehicle sensing device. s The distance D between RV and HV is calculated based on the position of RV. r The formula for calculating the weight is:
[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting vehicle behavior at intersections based on V2I, characterized in that, The system includes a prediction system comprising: a roadside sensing device (Sensor), an edge computing device (MEC), a roadside unit (RSU), and a vehicle (HV), wherein the roadside sensing device (Sensor), the edge computing device (MEC), the roadside unit (RSU), and the vehicle (HV) are connected in sequence. The prediction method includes the following steps: S1: The roadside sensing device will send the raw sensing data of the target vehicle (RV) and other surrounding targets to the MEC in real time; S2: MEC uses a deep learning-based multi-sensor fusion perception method to perceive target information of RV and other surrounding targets based on the raw perception data in S1; S3: MEC uses the RV's motion history and target information of other surrounding targets, combined with lane information in the high-precision map, and adopts a semantic map-based intersection prediction model to output the prior probability of each intention of the RV and predict the RV's driving intention. S4: MEC samples the trajectory of each RV intent based on the intersection map data, sampling multiple possible motion trajectories. Then, it calculates the cost of each motion trajectory by considering the cost function of acceleration and centripetal acceleration. Based on the magnitude of the cost, it selects the most reasonable trajectory for each intent. Finally, it calculates a posterior probability by estimating the likelihood of the cost function and combining it with the prior probability, thus obtaining the final trajectory prediction probability. S5: MEC will send RV target information, including intent prediction probability and motion prediction trajectory, to the autonomous vehicle HV via RSU; S6: The autonomous vehicle (HV) matches the RV target information sensed by the roadside equipment sent by the MEC with the multi-target information sensed by the autonomous vehicle's sensing equipment, filters out its own behavior prediction for the RV, and performs weighted fusion of the roadside prediction results and the vehicle-side prediction results based on the RV distance and tracking time to calculate the final RV behavior prediction result for use by the HV's decision planning module. In step S6, the specific method of weighted fusion is as follows: the driving intention of vehicle RV at the intersection includes four types: going straight, turning left, turning right, and making a U-turn. The output of behavior prediction is the probability of the four intentions and their corresponding driving trajectories. The predicted straight-ahead probability Prs, left-turn probability Prl, right-turn probability Prr, and U-turn probability Pru are given at the road end, where Prs + Prl + Prr + Pru = 1. The vehicle-side predictions for the straight-going probability Pvs, the straight-going probability Pvl, the right-turn probability Pvr, and the U-turn probability Pvu are given, where Pvs+Pvl+Pvr+Pvu=1. Let the weight at the road end be Wr and the weight at the vehicle end be Wv, where Wr + Wv = 1; We perform weighted fusion of data from the roadside and vehicle-side, and calculate the fused probability: The calculation of the vehicle-side weight Wv is related to the vehicle's sensing range and tracking duration, specifically: Based on the sensing range of the HV's autonomous vehicle sensing equipment and the RV's location, first determine whether the RV is within the effective sensing range of the autonomous vehicle sensing equipment: If RV is not within the perception range, the weight is calculated using the following formula: If the RV is within the sensing range, the maximum effective sensing distance of the vehicle sensing device is calculated as Ds based on its performance and installation location. The distance between the RV and the HV is calculated as Dr based on the RV's location. The weighting formula is as follows:
2. The method for predicting vehicle behavior at intersections based on V2I as described in claim 1, characterized in that, The autonomous vehicle (HV) is equipped with autonomous vehicle sensing devices and V2X communication devices (OBUs), which communicate with the vehicle unit (RSU) in real time.
3. The V2I-based vehicle behavior prediction method for intersections as described in claim 1, characterized in that, The roadside sensing equipment includes cameras and radar.
4. The V2I-based vehicle behavior prediction method for intersections as described in claim 1, characterized in that, In step S2, the target information includes, but is not limited to, position, speed, direction of travel, and size.
5. The V2I-based vehicle behavior prediction method for intersections as described in claim 1, characterized in that, In step S2, the specific implementation process of the deep learning multi-sensor fusion perception method is as follows: the target data perceived by multiple sensors is spatiotemporally aligned, the Hungarian algorithm is used to associate multiple targets, and the associated data is fused by Kalman filtering, so as to finally achieve target-level fusion of multi-sensor perception data.
6. The V2I-based vehicle behavior prediction method for intersections as described in claim 1, characterized in that, In step S6, the HV's vehicle perception devices include cameras, radar, and GPS.