Stereo Camera Object Detection via Parallax Density Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection devices struggle to accurately detect vehicles in scenes where the lamp of a preceding vehicle is turned off, such as at night or during braking, as they rely on lamp information, which is ineffective in these conditions.
Innovation Solution
An object detection device that uses a three-dimensional object detection unit to process images from stereo cameras, extracts combination candidates, determines sparse parallax regions, and matches regions between left and right images to identify vehicles regardless of lighting conditions by combining three-dimensional objects based on equal perspectives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If lamp information is used to detect vehicles, then detection performance in scenes with lamps turned on (night time, braking) is improved, but detection performance in scenes where lamps are turned off deteriorates
Solution Approach 1:
The patent segments the vehicle detection task into multiple independent detection channels: lamp-based detection and non-lamp-based detection (using body contours, size, shape). By dividing the detection problem into these segments, the system can selectively apply appropriate detection methods based on scene conditions, thereby maintaining high detection accuracy across both lamp-on and lamp-off scenarios
Solution Approach 2:
The patent dynamically switches between different detection strategies based on scene conditions. When lamps are detected, the system uses lamp information for detection; when lamps are not detected, it automatically transitions to using non-lamp features such as vehicle body contours and geometric properties. This dynamic adaptation ensures consistent detection performance across varying lighting conditions
2Device complexity
If three-dimensional objects are detected separately in left and right images, then detection process is simplified, but vehicle recognition accuracy deteriorates due to failure to recognize divided vehicle parts as a single vehicle
Solution Approach 1:
The patent merges detection results from left and right images by identifying and combining three-dimensional objects that represent different parts of the same vehicle. It uses spatial relationships, size constraints, and geometric consistency to determine when separate detected objects should be consolidated into a single vehicle entity, thereby improving recognition accuracy without significantly increasing process complexity
Solution Approach 2:
The patent introduces an intermediary processing stage that analyzes the relationship between detected three-dimensional objects from left and right images. This intermediary layer evaluates spatial consistency, size compatibility, and geometric relationships to determine whether separate detections correspond to the same vehicle, acting as a mediator that connects simple detection with accurate recognition
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
Enables stable recognition of vehicles in various scenes, improving detection accuracy by combining three-dimensional objects and determining sparse parallax regions, thus overcoming the limitations of lamp-based detection methods.
Implementation Method 1
a left image and a right image imaged by a left imaging unit and a right imaging unit
Data Source
AI summary
Provided is an object detection device capable of reliably recognizing a vehicle. This object detection device detects multiple three-dimensional objects from a left image and a right image captured with a left imaging unit and a right imaging unit (S103), extracts, as combination candidates from among the multiple three-dimensional objects, two three-dimensional objects which exist with an interval to the left/right, and determines whether a sparse parallax region, which is a region having a smaller parallax density than left/right regions, exists in an intermediate region between the two three-dimensional objects extracted as combination candidates. Then, the regions of two three-dimensional objects (for which it has been determined that a sparse parallax region exists in the intermediate region, and which have tentatively been identified as a single three-dimensional object) are extracted respectively from the left and right images and compared to each other, a determination is made regarding whether the perspective is the same, and when it is determined that the perspective is the same, the two three-dimensional objects are determined to be a single three-dimensional object.