Self-Driving Car Object Recognition via Reappearance-Based Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-driving car technologies face challenges in accurately recognizing objects on the road due to the infinite variety of features and limitations in processing capabilities, leading to errors in object recognition and distance estimation using cameras, radar, and lidar, which are costly and environmentally hazardous.
Innovation Solution
A method and device that improve object recognition by training a recognition model using training data from objects that disappear and reappear in images, incorporating active learning to enhance the model's performance, and applying filtering criteria to refine the training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object recognition is performed using camera images, then distance estimation can be achieved, but a lot of distance information is lost because the object in the real world is projected onto a two-dimensional image
Solution Approach 1:
The patent introduces an intermediary mechanism by detecting whether objects disappear and reappear in consecutive image frames. This disappearance-reappearance detection serves as a mediator to infer depth and distance information that is otherwise lost in 2D projections, allowing the system to estimate distances more accurately without adding physical depth sensors.
Solution Approach 2:
The system uses feedback from temporal changes in object visibility across multiple frames to continuously refine distance estimation. By monitoring when objects disappear from and reappear in the field of view, the system gains feedback about spatial relationships and uses this to improve ongoing distance measurements and object recognition accuracy.
2Reliability
If training data is collected from all recognized objects, then recognition model performance improves, but computational cost and processing time increase
Solution Approach 1:
The patent applies local quality by selectively focusing computational resources on specific objects that exhibit disappearance-reappearance patterns rather than processing all recognized objects uniformly. This targeted approach concentrates training efforts on edge cases and ambiguous situations where the recognition model benefits most, improving overall reliability without proportionally increasing computational cost.
Solution Approach 2:
The system performs partial action by collecting training data only from objects that meet specific criteria (disappearance and reappearance within defined time periods) rather than processing all possible training samples. This selective data collection provides sufficient training material to improve recognition accuracy while avoiding the excessive computational burden of processing complete datasets.
3Productivity
If filtering criteria are applied to training data, then training efficiency improves, but some potentially useful training information may be excluded
Solution Approach 1:
The filtering criteria are designed with feedback mechanisms that monitor recognition performance and adjust the selection of training data accordingly. The system evaluates whether excluded objects might provide valuable training information based on current model performance metrics, allowing dynamic adjustment of filtering thresholds to balance training efficiency with information retention.
Solution Approach 2:
The patent employs parameter changes by adjusting filtering thresholds and time period definitions based on environmental conditions, object types, and recognition performance. These dynamic parameter adjustments allow the system to optimize the balance between training data volume and training efficiency, adapting the filtering strictness to prevent information loss when conditions warrant more comprehensive data collection.
Data Source
AI summary
Provided is a method of improving an object recognition rate including recognizing a first object in a first image obtained while driving, detecting whether the recognized first object has disappeared for a preset time period and then reappeared in the first image, based on detecting that the first object has reappeared, calculating training data for the first object, and controlling such that a recognition model for recognizing an object included in an image is to be trained based on information based on the calculated training data.