LiDAR Dynamic Object Detection Without Texture-Rich Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection systems for autonomous vehicles rely on extensive training data and require adequate texture information, which is challenging to obtain with LiDAR data due to sensor viewpoint changes and scene occlusion, limiting their ability to accurately detect and track dynamic objects.
Innovation Solution
A machine learning model, such as a deep neural network (DNN), processes LiDAR range images and projection images to compute motion masks and vectors, using lightweight convolutional architectures that require minimal training data and are robust to noise and occlusion by incorporating multiple frames for input channels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional object detectors are used to identify objects across frames, then object tracking capability is provided, but the system requires extensive training data and prior knowledge of object types which limits detection of unlimited object types
Solution Approach 1:
The patent extracts the motion detection function from traditional object detectors by using optical flow algorithms that operate on raw LiDAR range images without requiring object classification. This separates motion analysis from object identification, allowing the system to detect motion of any object type without prior training data.
Solution Approach 2:
The system segments the detection task into two independent components: motion detection via optical flow on depth values, and object classification via object detectors. This segmentation allows motion detection to function independently without relying on object type priors, enabling detection of unlimited object types.
2Measurement precision
If optical flow approaches are used to find pixel-level flow field from LiDAR frames, then motion detection capability is provided, but the system requires adequate texture information which is challenging to obtain due to LiDAR sensor viewpoint changes and scene occlusion
Solution Approach 1:
The patent replaces traditional optical flow methods that rely on texture gradients with a depth-based optical flow approach using LiDAR range images. By computing motion from depth value changes rather than intensity gradients, the system achieves accurate motion detection without requiring texture information, overcoming LiDAR viewpoint changes and occlusion.
3Adaptability or versatility
If conventional systems combine object detectors and optical flow approaches, then both object identification and motion detection are provided, but the system does not overcome the shortcomings of requiring texture information and extensive training data
Solution Approach 1:
The patent transitions from 2D intensity-based optical flow to 3D depth-based motion analysis by utilizing the depth dimension inherent in LiDAR range images. This dimensional change enables motion detection to function independently of texture information, simplifying the system architecture while maintaining versatility.
Data Source
AI summary
In various examples, systems and methods of the present disclosure detect and/or track objects in an environment using projection images generated from LiDAR. For example, a machine learning model—such as a deep neural network (DNN)—may be used to compute a motion mask indicative of motion corresponding to points representing objects in an environment. Various input channels may be provided as input to the machine learning model to compute a motion mask. One or more comparison images may be generated based on comparing depth values projected from a current range image to a coordinate space of a previous range image to depth values of the previous range image. The machine learning model may use the one or more projection images, the one or more comparison images, and/or the one or more range images to compute a motion mask and/or a motion vector output representation.