Autonomous Vehicle Visual Perception Using Segmented Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles experience delays and reduced sensitivity in recognizing target objects due to the need to input visual perception images into deep learning models for feature extraction and recognition, leading to slower recognition speeds and increased risk of traffic accidents.
Innovation Solution
A visual perception method that involves obtaining visual perception images, inputting them into a first neural network model to eliminate redundant feature information, and then processing the multi-channel feature information into sub-neural network models, where each target recognition object corresponds to a specific sub-neural network model, reducing redundant feature processing time and improving recognition speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If visual perception images are input into deep learning models for feature extraction and recognition, then recognition accuracy is improved, but recognition speed deteriorates
Solution Approach 1:
The patent divides the single deep learning model into multiple sub-neural network models, each responsible for recognizing specific target objects. This segmentation allows parallel processing of different object types, improving recognition speed while maintaining accuracy through specialized processing for each object category
Solution Approach 2:
The patent performs preliminary feature extraction and eliminates redundant feature information before inputting data into sub-neural network models. This preliminary processing reduces the computational burden on subsequent recognition models, enabling faster recognition without sacrificing accuracy
2Measurement precision
If separate deep learning models are used for different target objects, then recognition accuracy is improved, but processing time increases
Solution Approach 1:
The system segments recognition tasks into specialized sub-neural network models for different object types (pedestrians, vehicles, cyclists, etc.). Each model is optimized for its specific target, improving accuracy while the parallel architecture reduces overall processing time compared to sequential processing
Solution Approach 2:
The patent merges multiple specialized sub-neural network models into a unified recognition system that processes multiple object types simultaneously. This combination maintains the accuracy benefits of specialized models while achieving parallel processing that reduces total processing time
Data Source
AI summary
The present disclosure provides a visual perception method, an apparatus, a device and a medium based on an autonomous vehicle, the method includes inputting an obtained first visual perception image collected by the autonomous vehicle into a first neural network model, recognizing multi-channel feature information of at least one target recognition object to be recognized, to eliminate redundant feature information in the first visual perception image; further, inputting the multi-channel feature information of the at least one target recognition object to be recognized into at least one sub-neural network model in a second neural network model respectively, to obtain at least one target recognition object; where there is a one to one correspondence between the target recognition object and the sub-neural network model. The present disclosure improves the speed of recognizing the target recognition object, thereby improving sensitivity of the autonomous vehicle and ensuring driving safety of the autonomous vehicle.