Point Cloud Obstacle Recognition Using Image-Based Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional obstacle recognition technologies in autonomous vehicles suffer from low accuracy.

Innovation Solution

Convert point cloud data into a first image using a neural network model for obstacle recognition, leveraging features learned in advance to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional obstacle recognition technologies are used, then the system is simpler, but the recognition accuracy is low

Engineering Contradiction:
Improveobstacle recognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network model with obstacle features before actual obstacle recognition. The model is trained in advance using training images and obstacle feature data, so that when real-time recognition is needed, the pre-trained model can quickly and accurately identify obstacles without requiring complex real-time processing of raw sensor data from scratch.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If neural network model is used for obstacle recognition, then recognition accuracy is improved, but computational burden increases

Engineering Contradiction:
Improveobstacle recognition accuracyVSAvoidcomputational energy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent reduces computational energy consumption by performing the computationally intensive neural network training in advance. The pre-trained model is then deployed for real-time obstacle recognition, which requires significantly less computational power compared to training the model in real-time. This separates the heavy computational task from the real-time operation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses training images as copies or representations of real obstacle scenarios to train the neural network model. These training images capture various obstacle features and conditions, allowing the model to learn from simulated data rather than requiring extensive real-time processing of actual sensor data during operation.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3872688B1Obstacle identification method and device, storage medium, and electronic device
Publication Date: 2026.04.22 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3872688B1 patent drawingFigure 1
  • EP3872688B1 patent drawingFigure 2~4
  • EP3872688B1 patent drawingFigure 5~9

AI summary

An obstacle identification method and device, a storage medium, and an electronic device. The method comprises: obtaining point cloud data in the automatic driving process of a target vehicle, wherein the point cloud data is obtained by scanning a driving environment of the target vehicle by a sensor on the target vehicle (S202); converting the point cloud data into a first image for representing the driving environment (S204); and identifying a first object located in the driving environment from the first image by means of a first neural network model (S206). The method solves the technical problem in the related art of the low identification accuracy of an obstacle.