Semantic Grid Map Fusion for Reliable Autonomous Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing grid maps for autonomous driving systems lack the accuracy and reliability needed for precise object detection, as they often rely on occupancy probabilities alone, and semantic segmentation models can introduce noise with multiple labels for a single object.
Innovation Solution
A method for fusing grid maps using a semantic segmentation model and a sensor fusion object detection model to generate a fused grid map with enhanced reliability by adjusting probabilities and determining representative labels based on the highest values, incorporating occupancy and non-occupancy probabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single semantic segmentation model is used to generate grid maps from point cloud data, then the system can process data efficiently, but the detection accuracy is limited and noise with multiple labels for single objects occurs
Solution Approach 1:
The patent combines multiple semantic segmentation models (first and second semantic segmentation models) to process point cloud data and generate grid maps. By merging the outputs of multiple models, the system achieves both efficient data processing and improved detection accuracy, reducing the noise and multiple label issues that occur with a single model.
2Device complexity
If occupancy probability information alone is used in grid maps, then the map structure remains simple, but the reliability and accuracy for autonomous driving are insufficient
Solution Approach 1:
The patent creates a composite grid map structure that integrates multiple types of information: occupancy probability, semantic labels, and confidence scores. This composite approach maintains a manageable grid map structure while significantly improving reliability for autonomous driving applications by combining different information layers.
3Loss of information
If semantic segmentation models assign multiple labels to single objects, then comprehensive object information is captured, but noise and detection reliability decrease
Solution Approach 1:
The patent applies local quality by assigning different confidence scores to different labels within the same object detection. Instead of treating all labels equally, the system evaluates the reliability of each label locally and uses confidence thresholds to filter out noisy labels, thereby maintaining comprehensive object information while improving overall detection reliability.
4Loss of time
If grid maps are generated without probability adjustment, then the generation process is faster, but the accuracy and reliability of the fused grid map are reduced
Solution Approach 1:
The patent performs preliminary probability adjustment on the outputs of individual semantic segmentation models before fusing them into the final grid map. By pre-adjusting confidence scores and filtering low-probability predictions beforehand, the system reduces the computational burden during fusion while maintaining high accuracy in the final fused grid map.
Data Source
AI summary
A method performed by an apparatus for controlling autonomous driving of a vehicle is introduced. The method may comprise generating, based on a segmentation model processing point cloud data, a first semantic grid map, generating, based on an object detection model, a second semantic grid map, adjusting a probability regarding whether occupancy exists for an element included in each grid of the first semantic grid map and the second semantic grid map, and generating a fused grid map by determining, as a representative label, at least one label corresponding to a highest value among final probabilities of the at least one label, wherein the final probabilities are determined based on whether the at least one label matches the element, outputting, based on the fused grid map, a signal, and controlling, based on the signal, autonomous driving of the vehicle.