Traffic Light Recognition Using Adaptive ROI Across Video Frames
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting tiny objects, such as traffic lights, in autonomous driving systems face challenges including decreased performance for large objects, reliance on accurate ground truth labeling, and limited adaptability to new situations.
Innovation Solution
A real-time tiny object detection methodology that does not require training, focusing on improving the inference method to enhance the detection rate for tiny objects by using prior information and detection results from previous images to adjust regions of interest.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training-based methods are used to improve detection performance for tiny objects, then detection rate for tiny objects is improved, but detection performance for large objects decreases
Solution Approach 1:
The patent divides the image processing into multiple stages: first performing detection on the entire image, then segmenting and enlarging specific regions containing tiny objects for secondary detection. This segmentation approach allows the system to maintain standard detection capabilities for large objects while applying enhanced processing specifically to tiny object regions, thus resolving the contradiction between improving tiny object detection and maintaining large object detection performance.
Solution Approach 2:
The patent applies different processing qualities to different regions of the image. Standard detection is applied to the entire image, while enhanced detection with enlargement is applied locally to specific regions containing tiny objects. This local quality enhancement allows improved tiny object detection without degrading the detection performance for large objects in other regions.
2Measurement precision
If training-based methods are used with accurate ground truth labeling, then detection performance is improved, but the requirement for additional labeling increases system complexity
Solution Approach 1:
The patent enables the detection system to improve its own performance through self-service mechanisms. By using detection results from previous frames and implementing adaptive region selection, the system automatically identifies and focuses on regions containing tiny objects without requiring external labeling input. The system learns from its own operational data to improve detection performance.
Solution Approach 2:
The patent implements feedback mechanisms by utilizing detection results from previous frames to inform current frame detection. The system feeds back detected object information across temporal frames, allowing it to adaptively adjust detection strategies and improve tiny object detection performance without requiring additional labeled training data.
3Measurement precision
If training-based methods are used to achieve robust detection, then detection accuracy is improved, but adaptability to new situations decreases
Solution Approach 1:
The patent implements dynamic adaptation by allowing the detection system to adjust its behavior based on real-time conditions. The region of interest selection and enlargement factors are dynamically determined based on detection results from previous frames and current image characteristics. This dynamic approach enables the system to adapt to new situations and varying driving conditions without requiring retraining.
Solution Approach 2:
The system uses feedback from continuous operation to adapt to new situations. By incorporating detection results from previous frames and using this information to guide current frame detection, the system continuously learns and adapts to new environmental conditions, object types, and scenarios without requiring additional labeled training data or model retraining.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
Disclosed is a method of detecting a traffic light with an object recognition model configured to recognize traffic lights. The method includes: obtaining an input image from a camera included in a vehicle, the input image among frames, including previous frames, captured by the camera; estimating, based on prior information about traffic light objects, a first region of interest (RoI) for the input image; determining a second RoI based on the first RoI and based on detection results of the previous frames, wherein the detection of results correspond to recognition results of recognizing traffic lights in the previous frames by the object recognition model; and recognizing, by the object recognition model, a traffic light in the input image, wherein the recognizing is based on the input image and the second RoI.