Event-Camera Object Detection for Time-to-Contact Under Variable Light
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous driving systems face challenges in accurately detecting adversarial objects and calculating time to contact, particularly under varying illumination conditions, which can lead to failure in object detection and segmentation.
Innovation Solution
The use of event cameras and dynamic vision sensors to filter events based on frequency analysis, combined with machine learning models for feature extraction and event filtering, enhances the detection of adversarial objects and calculates time to contact.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional visual perception systems are used for object detection, then the system complexity remains manageable, but the detection reliability for slow or static objects deteriorates leading to collisions
Solution Approach 1:
The patent segments the detection task by separating fast-moving object detection (using conventional frame-based cameras) from slow or static object detection (using event-based cameras with temporal difference processing). This segmentation allows each subsystem to be optimized for its specific detection scenario, improving overall reliability without requiring a complete system overhaul.
Solution Approach 2:
The patent introduces temporal difference processing as an intermediary mechanism that compares consecutive event-based camera frames to detect slow or static objects. This intermediary processing layer enhances detection capability for previously undetectable objects without significantly increasing overall system complexity.
2Measurement precision
If event-based camera systems with high temporal resolution are used, then the detection precision for fast approaching objects improves, but the processing complexity increases
Solution Approach 1:
The patent extracts only the essential temporal information from event-based camera data by computing temporal differences between consecutive frames. This extraction approach captures the most relevant information for detecting fast approaching objects while discarding redundant data, thereby maintaining detection precision without proportionally increasing processing complexity.
Solution Approach 2:
The patent performs preliminary temporal difference processing on event-based camera data before feeding it to object detection algorithms. This preliminary action pre-processes the high-resolution temporal data into a more manageable form, reducing the computational burden on subsequent processing stages while preserving critical detection information.
3Reliability
If configurable light sources are used to enhance detection of vulnerable road users, then the detection reliability improves, but the energy consumption increases
Solution Approach 1:
The patent employs periodic illumination from configurable light sources rather than continuous illumination. The light sources are activated in periodic cycles to enhance detection of vulnerable road users during critical periods, thereby improving detection reliability while significantly reducing overall energy consumption compared to continuous operation.
Solution Approach 2:
The patent applies configurable light sources locally to specific regions where vulnerable road users are detected or suspected to be present, rather than illuminating the entire environment. This localized approach concentrates energy where it is most needed for improving detection reliability while minimizing unnecessary energy consumption in other areas.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Improves the accuracy of object detection and time to contact calculation by filtering out adversarial objects, ensuring reliable autonomous driving performance under varying illumination conditions.
Implementation Method 1
an event based camera to detect a change in light intensity
Implementation Method 2
a convolutional neural network (CNN) to calculate the time to contact (TTC) of foreground objects
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for calculating a time to contact of an autonomous vehicle, the method comprising: obtaining a plurality of event data an image, wherein the event data is associated with a pixel associated with a change in light intensity; determining a reference signal frequency associated with a transmitted light; identifying a select event data from the plurality of event data, wherein the light frequency associated with the select event data is substantially the same as the reference signal frequency; determining an object based on the select event data, wherein the object is fully enclosed by a bounding box comprising coordinates of a rectangular border; calculating a distance between a set of coordinates of the bounding box closest to the autonomous vehicle and the autonomous vehicle; and calculating the time to contact between the set of coordinates of the bounding box and the autonomous vehicle.