Trajectory Planning Model Training With Road-Area Focus

Solution Overview

Problem

In low-speed automatic driving scenarios, neural networks face increased computation costs and resource consumption due to interference information like distant buildings, vehicles, and pedestrians in input images, which complicates accurate cruise path planning.

Innovation Solution

A method that involves obtaining images of the vehicle's environment, identifying road and non-road areas, determining initial trajectory points, and training a trajectory planning model based on the positional relationship between these points to focus on relevant data, reducing processing complexity and improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the neural network processes all objects in the input image including non-road objects, then the path planning accuracy may be improved by considering all environmental elements, but the computation cost and resource consumption increase significantly

Engineering Contradiction:

Improvepath planning accuracyVSAvoidcomputation cost

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts and removes non-road objects (distant buildings, vehicles, pedestrians) from the input image before processing by the neural network, retaining only road-related information. This extraction principle reduces the data volume processed while preserving the essential path planning information, thereby lowering computation cost without sacrificing planning accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different processing quality to different regions of the image: road areas are processed with high detail and precision, while non-road areas are either removed or processed with lower detail. This local quality differentiation ensures that computational resources are focused on the critical path planning regions rather than being wasted on irrelevant environmental elements.

Inventive Principle:

Principle #3Local quality

2Reliability

If the neural network processes all objects in the input image, then comprehensive environmental awareness is achieved, but the data processing time and resource consumption increase

Engineering Contradiction:

Improvepath planning reliabilityVSAvoiddata processing time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent extracts and removes non-essential non-road objects from the input image before neural network processing, retaining only road-related information critical for path planning. This extraction reduces the data volume processed while preserving the essential path planning information, thereby lowering computation cost without sacrificing planning accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different processing quality to different regions of the image: road areas are processed with high detail and precision, while non-road areas are either removed or processed with lower detail. This local quality differentiation ensures that computational resources are focused on the critical path planning regions rather than being wasted on irrelevant environmental elements.

Inventive Principle:

Principle #3Local quality

3Adaptability or versatility

If the neural network processes interference information from non-road objects, then the model may capture more environmental context, but the network capacity and computational resources are consumed unnecessarily

Engineering Contradiction:

Improveenvironmental context awarenessVSAvoidnetwork capacity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent extracts and removes non-essential non-road objects from the input image before neural network processing, retaining only road-related information critical for path planning. This extraction reduces the data volume processed while preserving the essential path planning information, thereby lowering computation cost without sacrificing planning accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different processing quality to different regions of the image: road areas are processed with high detail and precision, while non-road areas are either removed or processed with lower detail. This local quality differentiation ensures that computational resources are focused on the critical path planning regions rather than being wasted on irrelevant environmental elements.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS11940803B2Method, apparatus and computer storage medium for training trajectory planning model

Publication Date: 2024.03.26 BEIJING BAIDU NETCOM SCI & TECH CO LTD

AI summary

A method for training a trajectory planning model, an apparatus, and computer storage medium are provided. The method may include: obtaining an image of a physical environment in which a vehicle is located via at least one sensor of the vehicle, the image including multiple objects surrounding the vehicle; obtaining a feature chart indicating multiple initial trajectory points of the vehicle in the image from a trajectory planning model based on the image; identifying the image to determine in the image a first area associated with a road object in multiple objects and a second area associated with a non-road object in the multiple objects; determining a planning trajectory point based on positional relationship of the multiple initial trajectory points with respect to the first area and the second area; and training a trajectory planning model based on the planning track point and the actual trajectory point of the vehicle.