Range Sensor Prediction for Autonomous Obstacle Avoidance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Autonomous vehicles face challenges in predicting future sensory observations to create optimized trajectories and avoid obstacles, as existing methods are complex and may be sensitive to perturbations, especially when relying on image data.

Innovation Solution

A method using a temporal neural network that processes sequences of previous sensory observations and control actions to predict future sensory observations, incorporating a gated recurrent unit and multi-layer perceptrons, which enhances environmental perception and obstacle detection, making it less sensitive to perturbations and adaptable for active or passive range sensing devices.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing methods use image data for environmental perception, then obstacle detection capability is improved, but sensitivity to perturbations increases and computational complexity increases

Engineering Contradiction:
Improveobstacle detection capabilityVSAvoidsensitivity to perturbations
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent replaces image data processing with range sensing data (ultrasonic, lidar, radar) for environmental perception. This substitution of the sensing modality reduces sensitivity to image perturbations while maintaining obstacle detection capability, as range data provides direct distance measurements that are more robust to environmental variations.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the parameter space from image coordinates to range measurements. By transforming the perception task from processing pixel data to processing distance measurements, the system achieves better reliability while maintaining detection precision through the use of temporal neural networks on range sensing sequences.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If complex deep learning models are used for trajectory optimization, then prediction accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the prediction task into multiple temporal steps, where the temporal neural network predicts future range measurements step-by-step. This segmentation allows the use of a relatively simple recurrent architecture that processes sequences efficiently, achieving good prediction accuracy without requiring computationally intensive complex models.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The temporal neural network uses its own previous predictions as inputs for subsequent predictions, creating a self-service predictive mechanism. This recursive prediction approach allows the system to maintain accurate multi-step predictions while using a computationally efficient recurrent architecture rather than complex models.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3923200A1Prediction of future sensory observations of a distance ranging device
Publication Date: 2021.12.15 ELEKTROBIT AUTOMOTIVE GMBH
  • EP3923200A1 patent drawingFigure 1~2
  • EP3923200A1 patent drawingFigure 3~4
  • EP3923200A1 patent drawingFigure 5

AI summary

The present invention is related to a method, a computer program code, and an apparatus for predicting future sensory observations of a distance ranging device of an autonomous or semi-autonomous vehicle. The invention is further related to an autonomous driving controller using such a method or apparatus and to an autonomous or semi-autonomous vehicle comprising such an autonomous driving controller. In a first step, a sequence of previous sensory observations and a sequence of control actions are received (S1). The sequence of previous sensory observations and the sequence of control actions are then processed (S2) with a temporal neural network to generate a sequence of predicted future sensory observations. Finally, the sequence of predicted future sensory observations is output (S3) for further use.