LiDAR Feature Map Generation for Faster AI Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The conventional method of using LiDAR sensor data as input for artificial intelligence models in autonomous driving vehicles is inefficient due to the high computational load required for transforming three-dimensional point cloud data into tensor form, leading to prolonged object detection times and increased inefficiency in training and inference operations, especially when more channels are added for accuracy.
Innovation Solution
A method is introduced to generate an input feature map by combining channel-combined feature maps for distance and angle, pre-computed for a preset object detection area, using LiDAR sensor data in real-time, which includes generating first feature maps based on predefined indicators and combining them with pre-generated feature maps for distance and angle.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LiDAR sensor data is transformed into tensor form for AI model input, then object detection accuracy is improved, but computation time increases significantly
Solution Approach 1:
The patent pre-generates feature maps for distance and angle information before actual object detection. These feature maps are created in advance and stored, so when real-time detection is needed, the system only needs to combine them with current LiDAR data rather than computing everything from scratch. This preliminary preparation significantly reduces real-time computation time while maintaining detection accuracy.
Solution Approach 2:
The patent divides the complex transformation process into separate feature map generations for different indicators (distance, angle, etc.). Each feature map is generated independently and then combined through channel combination. This segmentation allows parallel processing and optimization of individual components, reducing overall computation time while preserving the accuracy benefits of comprehensive feature extraction.
2Measurement precision
If multiple channels are added to input feature map for more accurate results, then detection precision is improved, but training and inference efficiency deteriorates
Solution Approach 1:
The system pre-computes and stores feature maps for multiple indicators (distance, angle, and other spatial features) that would otherwise require heavy real-time computation. By having these feature maps ready in advance, the system can use multiple channels for precise detection without the computational burden of generating all features in real-time, thus maintaining both precision and efficiency.
Solution Approach 2:
The patent creates multiple feature map channels that represent different aspects of the environment (distance, angle, etc.). These are essentially copies of processed spatial information organized into different channels. Using these pre-processed copies instead of raw data allows multiple channels to be used for accurate detection while avoiding the computational cost of processing equivalent information multiple times during training and inference.
Data Source
AI summary
Provided are a method, an apparatus, and a computer program for generating an input feature map of an artificial intelligence model to detect an object based on light detection and ranging (LiDAR). A method of generating an input feature map of an artificial intelligence model to detect an object based on LiDAR, which is to be performed by a computing apparatus, may include collecting LiDAR sensor data in the form of a three-dimensional point cloud for a predetermined area, generating one or more first feature maps by processing the collected LiDAR sensor data based on one or more predefined indicators, and generating the input feature map of the artificial intelligence model by combining, as an independent channel, each of the one or more generated first feature maps with one of the second feature maps previously generated based on a distance and an angle.