Bird's-Eye 3D Object Learning for Low-Complexity Obstacle Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing obstacle detection systems for mobile objects face increased system costs due to complex hardware configurations when using multiple ranging sensors, while relying solely on cameras requires extensive training data for robustness.
Innovation Solution
A learning method and device that associate annotations with bird's-eye view images to generate a trained model capable of detecting three-dimensional objects using a smaller amount of data, allowing for efficient obstacle detection without complicating the hardware configuration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple ranging sensors are used for obstacle detection, then detection accuracy is improved, but hardware complexity and system costs increase
Solution Approach 1:
The patent combines multiple sensor types (camera, LIDAR, ultrasonic sensor) into a unified sensing system that processes information through a single trained model. This merging approach maintains the detection advantages of multiple sensors while simplifying the overall system architecture and reducing hardware complexity.
Solution Approach 2:
The trained model serves as a universal processing unit that handles detection tasks for all sensor types. Instead of requiring separate processing systems for each sensor, the multi-functional model processes inputs from cameras, LIDAR, and ultrasonic sensors through a single architecture, reducing system complexity.
2Device complexity
If only a camera is used to simplify hardware configuration, then system costs are reduced, but the amount of training data required increases significantly
Solution Approach 1:
The patent uses a composite sensing approach that combines data from multiple sensor modalities (visual data from camera, depth data from LIDAR, acoustic data from ultrasonic sensors). This composite input allows the system to achieve robust obstacle detection with less training data compared to a camera-only system, as each sensor type provides complementary information that reduces the learning burden.
3Reliability
If multiple sensor types are integrated, then robustness for various scenes is improved, but system costs increase
Solution Approach 1:
The trained model acts as an intermediary that receives processed inputs from multiple sensor types and produces unified obstacle detection outputs. This intermediary layer allows the system to leverage the robustness advantages of multiple sensors while presenting a simplified interface to the control system, effectively managing the complexity-reliability trade-off.
Data Source
AI summary
A learning method includes steps of associating, with an extended area in a bird's-eye view image, an annotation indicating that the extended area is a three-dimensional object; and generating, based on training data in which a bird's-eye view image is associated with an annotated bird's-eye view image obtained by assigning an annotation to a three-dimensional object in the bird's-eye view image, a trained model by learning parameters of a machine learning model so that the trained model receives input of a bird's-eye view image to output a three-dimensional object in the bird's-eye view image.