Obstacle-to-Path Assignment Using Rasterized DNN Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for obstacle to path assignment in autonomous systems require significant computing resources, processing time, and are inaccurate due to the projection of 3D lane information to 2D space, especially in scenarios with high curvature or distant objects, leading to increased complexity and reduced real-time capability.
Innovation Solution
The use of a deep neural network (DNN) that incorporates rasterized path and object representations, along with sensor data, to perform binary classification and filter noise, reducing the need for separate path and object detection, and employing data augmentation to enhance training data accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional post-processing methods are used to determine object to lane assignments by projecting 3D lane locations to 2D space, then lane detection can be performed separately from object detection, but this requires significant computing resources, additional processing time, and results in inaccurate assignments especially in scenarios with high curvature or distant objects
Solution Approach 1:
The patent combines lane detection and object detection into a single unified neural network model that processes sensor data and directly outputs object-to-lane assignments. This eliminates the need for separate detection tasks and post-processing projection operations, thereby improving assignment accuracy while reducing computational overhead and processing time.
2Adaptability or versatility
If separate detection of objects in 2D space and lanes in 3D space is performed prior to determining intersections, then the detection tasks can be independently optimized, but this requires significant computing resources and additional processing time which presents limitations to real-time autonomous driving systems
Solution Approach 1:
The patent merges separate object detection and lane detection processes into a single integrated neural network that performs both tasks simultaneously. The unified model takes sensor data as input and directly produces object-to-lane assignment outputs, eliminating the need for independent optimization of separate detection tasks and significantly improving real-time processing speed.
3Device complexity
If the projection from 3D to 2D is performed with a flat world assumption, then the processing can be simplified, but the resulting projection is inaccurate because the path includes curves, dips, hills, and pitch
Solution Approach 1:
The patent replaces the mechanical projection approach (which requires complex 3D-to-2D coordinate transformations and flat world assumptions) with a neural network-based system. The neural network learns to directly map sensor data to object-to-lane assignments, naturally handling curved paths, dips, hills, and pitch without requiring explicit geometric projections, thereby maintaining simplicity while improving accuracy.
4Adaptability or versatility
If a deep neural network is trained to handle different scenarios that are less common and complex, then the DNN can intrinsically interpret lane and object information from input images, but this increases the compute requirements and complexity of the DNN as well as lead to less accurate assignments in scenarios where objects are far away and/or on paths with high curvature
Solution Approach 1:
The patent enhances the neural network's ability to handle diverse scenarios by incorporating additional input dimensions beyond just the input image. The model processes sensor data that includes explicit lane and object location information, allowing the network to make accurate assignments even for distant objects or objects on high-curvature paths without requiring excessive network complexity. This multi-dimensional input approach enables the model to learn from various scenarios effectively.
Data Source
AI summary
In various examples, one or more output channels of a deep neural network (DNN) may be used to determine assignments of obstacles to paths. To increase the accuracy of the DNN, the input to the DNN may include an input image, one or more representations of path locations, and/or one or more representations of obstacle locations. The system may thus repurpose previously computed information—e.g., obstacle locations, path locations, etc.—from other operations of the system, and use them to generate more detailed inputs for the DNN to increase accuracy of the obstacle to path assignments. Once the output channels are computed using the DNN, computed bounding shapes for the objects may be compared to the outputs to determine the path assignments for each object.