Vehicle Lamp Recognition with Recurrent Encoder-Decoder Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for recognizing activated vehicle lamps in driver assistance systems and semi-automated driving require multiple chronologically consecutive frames and are computationally intensive, often failing to reliably classify lamps in a single image due to isolated processing.
Innovation Solution
A method utilizing a recurrent encoder network (ERNN) and decoder network (DRNN) processes multiple image recordings in a sequential manner, breaking down the task into subtasks with limited neuron counts, allowing reliable lamp classification even on resource-constrained hardware, and incorporating contextual information from surrounding images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple chronologically consecutive frames are used for lamp recognition, then recognition reliability is improved, but processing time and computational complexity increase
Solution Approach 1:
The patent segments the lamp recognition task by dividing the vehicle image into multiple candidate areas (e.g., left side, right side, center) and processing each area independently through separate passes of the ERNN. This allows the system to achieve reliable recognition by examining multiple spatial regions simultaneously rather than requiring multiple temporal frames, thus improving reliability while reducing processing time.
2Measurement precision
If a comprehensive neural network is used for lamp recognition, then recognition accuracy is improved, but computational resources and hardware requirements increase
Solution Approach 1:
The patent applies segmentation by dividing the neural network processing into multiple passes, where each pass handles a specific candidate area. Each pass uses a limited ERNN with a small number of neurons (e.g., 10 neurons), rather than employing a single comprehensive network. This segmented approach achieves high classification accuracy by systematically examining all candidate areas while keeping individual network complexity low, thereby reducing hardware requirements.
3Productivity
If isolated processing of individual lamp areas is used, then processing speed is improved, but recognition reliability decreases due to lack of contextual information
Solution Approach 1:
The patent merges the processing of multiple candidate areas by executing the ERNN multiple passes, where each pass processes a different candidate area while utilizing the output from previous passes. This combining approach allows the system to process multiple areas efficiently (maintaining productivity) while incorporating contextual information from all areas into the final recognition decision, thereby improving reliability.
4Area of stationary object
If reflective surfaces like license plates are processed, then comprehensive area coverage is improved, but false positive rates increase due to light reflections
Solution Approach 1:
The patent segments the vehicle surface into distinct candidate areas (e.g., separating the license plate area from actual lamp areas). By processing each segment independently through multiple passes of the ERNN, the system can apply different processing strategies to reflective surfaces versus actual lamps, reducing false positives while maintaining comprehensive coverage.
Solution Approach 2:
The patent employs feedback mechanisms where the output of each ERNN pass is supplied as input to subsequent passes. This feedback loop allows the system to iteratively refine its classification, particularly for challenging areas like reflective surfaces. The feedback from previous passes enables the DRNN to learn patterns that distinguish actual lamps from reflective surfaces, improving detection accuracy while maintaining area coverage.
Data Source
AI summary
A method for recognizing which lamps at a vehicle are activated. The method includes: providing multiple image recordings of candidate areas at the vehicle in which an activated lamp is presumed; converting the image recordings into an intermediate product by executing a recurrent encoder network (ERNN), the output of at least one pass of the ERNN is supplied as input to a further pass of the ERNN, and different image recordings of candidate areas are supplied as input to different passes of the ERNN; assignments of the image recordings of candidate areas are ascertained to classes which represent specific lamps of the vehicle from the intermediate product by executing a recurrent decoder network (DRNN) multiple times, the output of at least one pass of the DRNN is supplied as input to a further pass of the DRNN.