Vehicle Object Detection Using Intermediate Representation Data Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-driving vehicle systems struggle to detect objects of unknown types, leading to undetected objects due to the failure of machine learning units to associate extracted features with known object types.
Innovation Solution
A method and system where a first vehicle uses an additional sensor to detect objects and verify the presence of unknown objects, extracting representation data from the machine learning unit's intermediate processing elements. These data are then transferred to a second vehicle, which uses a detection logic to compare local representation data with the received reference data, generating an auxiliary detection signal if a predefined similarity criterion is met.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the machine learning unit is trained to recognize all possible object types, then the detection reliability for unknown objects improves, but the resource consumption and system complexity increases significantly
Solution Approach 1:
The patent segments the object detection task into two parts: a machine learning unit for known object types and a separate detection logic for unknown objects. The detection logic uses intermediate representation data from the neural network to identify unknown objects without requiring the entire system to be retrained for every new object type, thus improving reliability while managing complexity
Solution Approach 2:
The patent introduces intermediate representation data as an intermediary between the machine learning unit and the detection logic. This representation data captures essential features of objects without committing to specific object type classifications, enabling the system to detect unknown objects by comparing representation data against stored patterns without requiring comprehensive training on all possible object types
2Reliability
If the machine learning unit is trained to recognize all possible object types, then the detection reliability improves, but the training resources and time required increases
Solution Approach 1:
The patent performs preliminary extraction of intermediate representation data during the normal operation of the machine learning unit, before unknown objects need to be detected. This representation data is stored and can be quickly compared against new objects without requiring time-consuming retraining of the entire neural network, thus improving detection reliability while minimizing training time losses
3Reliability
If additional sensors are deployed in the second vehicle to detect unknown objects, then the detection capability improves, but the device complexity and cost increases
Solution Approach 1:
The patent creates a virtual copy of the detection capability by transferring intermediate representation data from the first vehicle to the second vehicle. Instead of physically copying sensors or the entire machine learning unit, the essential detection knowledge is replicated through data transfer, enabling the second vehicle to detect unknown objects without duplicating the full sensor suite or computational infrastructure
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention is concerned with detecting an object (21) in surroundings (17) of a second vehicle (12), wherein a first vehicle (11) and the second vehicle (12) each comprise a machine learning unit (15) for recognizing an object (21) type of the object on the basis of respective sensor data (18). The invention is characterized in that the first vehicle (11) uses at least one additional sensor (27) for detecting the object (21) independently from the machine learning unit (15) and sends predefined representation data (33) of at least one intermediate processing element (25) of the machine learning unit (15) as reference data (34) to the machine learning unit (15) of the second vehicle (12). The machine learning unit (15) of the second vehicle (12) operates a detection logic (30) which compares its own representation data (33') of at least one intermediate processing element (25) with the reference data (34).