Road User Behavior Prediction With Lane-Segment Neural Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to accurately predict the near-future behavior of road users, such as driving intentions and trajectories, which is crucial for autonomous driving systems (ADS) due to intrinsic uncertainties in human driving behavior.
Innovation Solution
An assessment system that utilizes a combination of neural networks to process vehicle and road user states, lane segment data, and historical information to predict near-future intentions and trajectories by encoding and spatially-temporally modeling road user interactions within a selected coordinate system, incorporating multi-head and self-attention operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional prediction methods are used for road user behavior, then the system complexity is low, but the prediction reliability is insufficient due to intrinsic uncertainties in human driving behavior
Solution Approach 1:
The system segments the prediction task into multiple specialized neural networks: a first neural network for spatial encoding of road user states, a second neural network for temporal encoding using historical data, and a third neural network for behavior prediction. This segmentation allows each network to focus on specific aspects of the complex prediction problem, improving overall reliability while managing system complexity through modular architecture.
Solution Approach 2:
The system transforms the prediction problem from a single-dimension approach to a multi-dimensional framework by incorporating both spatial dimensions (through coordinate system transformation and encoding) and temporal dimensions (through historical data processing). This dimensional expansion enables the system to capture complex patterns in road user behavior that single-dimension methods miss, thereby improving prediction reliability.
2Measurement precision
If detailed spatial and temporal modeling is performed using multiple neural networks, then the prediction accuracy improves, but the computational complexity increases
Solution Approach 1:
The computational task is divided into three distinct neural networks, each handling a specific aspect: spatial encoding, temporal encoding, and behavior prediction. This segmentation reduces the computational burden on any single network while maintaining high overall accuracy through the coordinated work of multiple specialized networks.
Solution Approach 2:
The system performs preliminary spatial and temporal encoding of road user states before the actual behavior prediction. By pre-processing the data through dedicated encoding networks, the system prepares refined inputs for the final prediction network, improving accuracy while distributing computational complexity across multiple stages rather than concentrating it all in one step.
3Reliability
If the system processes multiple data dimensions including historical data and spatial coordinates, then the prediction reliability improves, but the data processing time increases
Solution Approach 1:
The data processing pipeline is segmented into parallel streams: spatial encoding processes coordinate information while temporal encoding processes historical sequences. This segmentation allows simultaneous processing of multiple data dimensions, reducing overall processing time while maintaining the comprehensive analysis needed for reliable predictions.
Solution Approach 2:
The system performs preliminary encoding of spatial and temporal data before the final prediction step. By pre-processing and organizing the multi-dimensional data into structured representations, the system reduces the computational burden of the final prediction, thereby decreasing total processing time while preserving prediction reliability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to a method performed by an assessment system (1) for predicting near-future behavior of road users in surroundings of a vehicle (2). The assessment system obtains (1001) a state of the vehicle, which vehicle state comprises at least a pose of the vehicle in view of a digital map (22); obtains (1002) from the digital map, map data associated with the vehicle pose, wherein the map data is transformed to a selected coordinate system; obtains (1003) states of road users (4) in the vehicle's surroundings; and identifies (1004) by assessing the road users states in view of the map data in the selected coordinate system, lane segments (4) void from road users (3). Furthermore, the assessment system encodes (1005) network input (5) comprising data associated with the lane segments, data associated with the road users states and historic data associated with the road users states and/or lane segments, into respective states-related data (61) associated with the road users states and segment-related data (62) associated with dynamic start and end boundaries of the lane segments. Moreover, the assessment system feeds (1006) respective states-related data and segment-related data to one or more neural networks configured to encode the road users states in view of the lane segments spatially and temporally, and output spatial- and temporal-processed respective states-related data (81) and segment-related data (82). The assessment system further feeds (1007) output spatial- and temporal-processed respective states-related data (811) and segment-related data (821) associated with one or more of the road users - referred to as target road user(s) (31) - to at least a first behavior-predicting neural network (9) configured to, by processing said data, predict and output data (90) indicating predicted near-future behavior of the target road user(s) in view of one or more of the lane segments. The disclosure also relates to an assessment system in accordance with the foregoing, a vehicle comprising such an assessment system, and a respective corresponding computer program product and non-volatile computer readable storage medium.