Autonomous Vehicle Trajectory Planning With Cascaded ML Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Self-driving cars face challenges in predicting and determining trajectories of other road users due to the computational intensity of determining object-specific features, which can lead to increased risk of collisions.
Innovation Solution
A computer-implemented method and system that utilize a cascade of machine-learning models, where each model generates increasingly accurate predictions based on the availability of computational resources, allowing for efficient and accurate motion planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the processor determines all object-specific features for each object in the surroundings using machine-learning algorithms, then the accuracy of trajectory prediction and collision risk assessment is improved, but the computational time and processing load increase significantly
Solution Approach 1:
The patent segments the object detection process into multiple sequential machine-learning models with increasing accuracy. The first model provides quick basic detection, while subsequent models progressively refine the detection accuracy for critical objects, allowing the system to allocate computational resources dynamically based on risk assessment rather than processing all objects with maximum accuracy simultaneously
Solution Approach 2:
The patent applies different levels of detection quality to different objects based on their relevance to the autonomous vehicle's trajectory and collision risk. High-priority objects (those near the vehicle's path or exhibiting risky behavior) receive more intensive processing with multiple ML models, while low-priority objects use faster, less accurate models, optimizing the balance between overall accuracy and computational efficiency
2Measurement precision
If the system processes sensed data through multiple sequential machine-learning models to improve prediction accuracy, then the object detection precision is improved, but the device complexity and processing overhead increase
Solution Approach 1:
The patent implements a dynamic processing architecture where the number and type of ML models applied to each object are not fixed but adapt based on real-time conditions. The system dynamically adjusts the detection pipeline complexity based on factors such as object priority, computational resource availability, and time constraints, allowing the system to optimize performance without requiring the maximum complexity to be always active
Solution Approach 2:
The patent performs preliminary processing with faster, less accurate ML models before applying more complex models. This staged approach allows the system to quickly filter and prioritize objects, performing intensive processing only on those that require it, thereby reducing the overall computational burden and system complexity compared to applying high-accuracy models to all objects
3Productivity
If the autonomous vehicle uses minimal computational resources for trajectory planning, then the operational efficiency is improved, but the ability to predict other road users' trajectories and avoid collisions deteriorates
Solution Approach 1:
The patent changes key parameters such as detection threshold, model complexity level, and processing depth based on real-time operational context. When collision risk is high or computational resources are abundant, the system increases detection accuracy and processing depth. When resources are constrained or risk is low, it reduces processing intensity, thereby maintaining operational efficiency while adapting reliability to match actual needs
Data Source
AI summary
A method and a system for planning motion of a Self-Driving Car (SDC) are provided. The method comprises: receiving sensed data representative of the surroundings of the SDC; feeding the sensed data to a plurality of machine-learning (ML) models, the feeding comprising: feeding the sensed data to a first ML model to generate a first prediction for a given object in the surroundings of the SDC; in response to a time for generating the first prediction being higher than a predetermined threshold: planning the motion of the SDC based on the first prediction; and in response to the time for generating the first prediction being lower than the first predetermined time period threshold: feeding the first prediction along with the sensed data to a second ML model to generate a second prediction for the given object; and planning the motion of the SDC based on the second prediction.