CNN Training for Unknown Object Detection in Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving and automated valet parking systems struggle to detect unknown objects, as they rely on trained neural networks that are not equipped to recognize objects outside their predefined classification, potentially leading to dangerous situations when encountering rare or unfamiliar objects.
Innovation Solution
A training method for convolutional neural networks (CNNs) that includes a class for unknown objects and uses images with no objects to enhance foreground-background classification, allowing the network to detect unknown objects by learning to differentiate between background and objects, even if the object has not been previously trained for.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are trained with predefined object classifications, then detection reliability for known objects is improved, but the ability to detect unknown objects deteriorates
Solution Approach 1:
The patent inverts the traditional object detection approach by training the neural network to detect the absence of known objects (background) rather than focusing solely on detecting specific object classes. This inversion enables the system to identify unknown objects as deviations from the expected background pattern, resolving the contradiction between reliable known object detection and unknown object adaptability
Solution Approach 2:
The patent changes the detection parameter from object-specific features to background-subtraction features. By modifying what the neural network detects (from object presence to object absence relative to background), the system achieves both reliable known object detection through training and unknown object detection through anomaly identification
2Adaptability or versatility
If motion-based detection methods are used for unknown objects, then unknown object detection capability is improved, but requirement for video sequences over time increases system complexity
Solution Approach 1:
The patent extracts the temporal dimension from the detection process by using background subtraction on individual frames rather than requiring analysis of video sequences over time. This extraction of the essential detection function from its complex temporal context reduces system complexity while maintaining unknown object detection capability
Solution Approach 2:
The patent creates a background model (a copy of the expected scene without objects) and compares current frames against this model. This copying approach enables unknown object detection without requiring complex temporal analysis of multiple frames, reducing system complexity
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Abstract (in connection with Fig. 1) The present invention concerns a training method, a method, a software program and a system for the detection of unknown objects, in particular for autonomous driving, autonomous manoeuvring or automated valet parking (AVP). The training method for the detection of unknown objects in a real-word environment, in particular for autonomous driving, autonomous manoeuvring or automated valet parking, comprises the following steps: a) collecting an image dataset including both images that contain objects to be detected as well as images that do not contain any objects to be detected, including images with random unknown objects, b) labelling the collected images, wherein both known and unknown objects are labelled in all the images of the dataset according to their respective class or classes, the respective class or classes including or consisting of a class for unknown objects, and their location in the respective image, and c) training a convolutional neural network (CNN) with the collected and labelled image dataset.