Multi-Task Object Detection With Optimal Fusion Paths for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional multi-task detection in vehicle autonomous driving using neural networks results in reduced detection efficiency and accuracy due to independently assigned backbone networks for each task.
Innovation Solution
A multi-task object detection method involving a multi-scale feature extraction network, a feature pyramid unit, and multiple detection heads, with optimal fusion paths trained through a training and verification process to enhance detection accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a shared neural network is used for multi-task detection, then device complexity is reduced, but detection accuracy and efficiency significantly decrease
Solution Approach 1:
The patent divides the detection system into separate task-specific detection heads (object detection head, lane detection head, drivable area segmentation head) while sharing the backbone feature extraction network. This segmentation allows each task to be optimized independently with task-specific parameters and loss functions, resolving the contradiction by maintaining low overall complexity through sharing while achieving high accuracy through specialization.
Solution Approach 2:
The backbone network serves as a universal feature extractor that provides multi-scale features to all detection heads simultaneously. This multi-functional design allows a single network component to serve multiple purposes, reducing overall device complexity while maintaining the ability to perform multiple detection tasks with high accuracy through task-specific optimization in the detection heads.
2Measurement precision
If independently assigned backbone networks are used for each task, then detection accuracy is improved, but multi-task detection efficiency decreases
Solution Approach 1:
The patent merges multiple task-specific detection processes by implementing a shared backbone network that extracts features once and reuses them across all detection tasks. The multi-scale feature fusion network combines features from different layers, and multiple detection heads process tasks simultaneously using the same feature representations, thereby improving efficiency while maintaining accuracy through task-specific optimization.
Solution Approach 2:
The patent introduces a multi-scale feature dimension by extracting and fusing features at multiple levels of the network. This dimensional approach allows the system to capture both fine-grained and coarse-grained features simultaneously, enabling accurate detection across different task types while processing efficiency is maintained through the shared backbone structure.
3Device complexity
If conventional shared neural networks are used, then device complexity is low, but detection accuracy cannot meet actual requirements
Solution Approach 1:
The patent applies local quality by making the detection heads task-specific with optimized parameters, loss functions, and processing pipelines tailored to each detection task's requirements. While the backbone remains general-purpose for efficiency, each detection head is locally optimized for its specific task (object detection, lane detection, or drivable area segmentation), ensuring high reliability without excessive overall complexity.
Data Source
Figure 1~2
Figure 3~4a
Figure 4b~4d
AI summary
The disclosure relates to the field of autonomous driving technologies, and specifically provides a multi-task object detection method, an electronic device, a medium, and a vehicle, to solve the technical problem of low detection accuracy or poor detection effect of an existing multi-task detection method. For this purpose, the multi-task object detection method of the disclosure includes: obtaining images captured by a vehicle-mounted sensor; inputting the images into a multi-scale feature extraction network to extract multi-scale features; inputting the multi-scale features into a multi-scale feature fusion network to obtain fused features, where the multi-scale feature fusion network includes multiple optimal fusion paths, and each optimal fusion path corresponds to one of multiple tasks; and inputting, into a corresponding detection head, the fused features output from each optimal fusion path, to obtain a detection result, where each detection head is capable of detecting one of the multiple tasks. In this way, the accuracy of multi-task object detection is improved.