Multi-Task Target Detection With Staged Backbone Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The reliability of multi-task target detection in autonomous driving is affected by high task conflicts, low convergence speed, and poor model training effects due to the use of multi-source data sets with varying task distributions and large numbers of tasks.

Innovation Solution

A multi-task target detection method involving a step-by-step training process, utilizing a shared main backbone network, branch backbone networks for different task groups, and individual head networks, with staged training on selected and fixed parameters to optimize model parameters and reduce task conflicts.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a multi-task learning model uses a single backbone network to extract depth features for multiple tasks, then computing resources are reduced and delay is minimized, but task conflicts increase and training reliability deteriorates when using multi-source data sets with large numbers of tasks

Engineering Contradiction:
Improvecomputing resource efficiencyVSAvoidtraining reliability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent divides the backbone network into a first backbone network and a second backbone network. The first backbone network processes common features shared across multiple tasks, while the second backbone network processes task-specific features. This segmentation reduces task conflicts by isolating different task requirements into separate processing paths while maintaining computational efficiency through shared common features.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If multi-source data sets with large numbers of tasks are used for training, then the model's adaptability improves, but task conflicts increase and convergence speed decreases

Engineering Contradiction:
Improvemodel adaptabilityVSAvoidconvergence speed
Core Design Contradiction:
Adaptability or versatilityVSSpeed

Solution Approach 1:

The patent segments the training process into two stages: first-stage training using only the first backbone network with a subset of tasks to establish common features, and second-stage training using both backbone networks with all tasks to optimize task-specific features. This segmentation enables the model to handle large numbers of tasks from multi-source data sets while maintaining fast convergence by progressively building knowledge rather than attempting to learn all tasks simultaneously.

Inventive Principle:
Principle #1Segmentation

3Adaptability or versatility

If all tasks are trained simultaneously using multi-source data sets, then the model's versatility improves, but training time increases and convergence speed decreases

Engineering Contradiction:
Improvemodel versatilityVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by conducting first-stage training before second-stage training. In the first stage, the model is trained on a subset of tasks using only the first backbone network to establish common feature representations. This preliminary training prepares the model for subsequent second-stage training where all tasks are processed using both backbone networks. This preliminary action significantly reduces the time required for full training by pre-establishing common features that can be reused across all tasks.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4276696B1Method, device, autonomous driving system and storage medium for multi-task target detection
Publication Date: 2025.08.20 ANHUI NIO AUTONOMOUS DRIVING TECH CO LTD
  • EP4276696B1 patent drawingFigure 1~2
  • EP4276696B1 patent drawingFigure 3~4

AI summary

The disclosure provides a multi-task target detection method and device, an autonomous driving system, and a storage medium. The method includes: selecting some of tasks in each task group to perform first-stage training on a model to be trained, to obtain an initial trained model, which prevents task conflicts, thereby increasing the convergence speed; fixing parameters of a main backbone network in the initial trained model, and performing second-stage training on the initial trained model by using all tasks in each task group, to obtain an intermediate trained model, which implements the training of different tasks in a uni-source data set or a multi-source data set without mutual interference, thereby shortening the training time; and releasing parameters of the main backbone network in the intermediate trained model, performing third-stage training on the intermediate trained model by using all tasks in all task groups to obtain a multi-task target detection model, inputting an image under detection that contains a target detection object into the multi-task target detection model, and detecting and recognizing the target detection object, which improves the reliability of multi-task target detection.