Obstacle-to-Path Assignment Using DNN Path and Object Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for obstacle to path assignment in autonomous driving systems require significant computing resources and time, leading to inaccurate and noisy outputs due to the projection of 3D lane data to 2D space and the complexity of deep neural networks (DNNs) in interpreting lane and object information from input images.
Innovation Solution
The system uses one or more output channels of a deep neural network (DNN) to determine obstacle to path assignments by inputting an image along with explicit representations of path and obstacle locations, such as rasterized images indicating lane shape and object locations, to increase the accuracy of obstacle to path assignments.
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 data to 2D space, then lane detection can be performed as a separate task, but computing resources and processing time are significantly increased
Solution Approach 1:
The patent combines lane detection and object detection into a single unified neural network model that processes sensor data simultaneously. The network outputs both lane information and object information in one processing pass, eliminating the need for separate detection tasks and post-processing steps. This merging approach directly reduces computing resources and processing time while maintaining the ability to perform both detection functions.
Solution Approach 2:
The unified neural network performs preliminary processing of sensor data to extract both lane and object information in advance, before any assignment operations are needed. By pre-computing both detection results simultaneously in a single forward pass, the system eliminates subsequent post-processing steps that would otherwise be required to synchronize and align separate detection outputs.
2Ease of operation
If 3D lane locations are projected to 2D space for comparison with 2D bounding shapes, then object to lane assignment can be determined, but accuracy decreases due to flat world assumption and path curvature
Solution Approach 1:
Instead of projecting 3D lane data to 2D space, the patent maintains lane information in 3D space throughout the neural network processing. The unified network operates in the original sensor data dimensionality, comparing object positions directly with lane positions in 3D space. This approach preserves spatial relationships and eliminates projection errors caused by flat world assumptions and path curvature.
3Extent of automation
If a deep neural network is trained to interpret lane and object information solely from input images, then the system can perform obstacle to path assignment, but compute requirements and DNN complexity increase
Solution Approach 1:
The patent designs a universal neural network that performs multiple functions simultaneously: detecting lanes, detecting objects, and determining assignments between them. This single multi-functional model replaces what would otherwise require separate specialized networks for each task. The unified architecture reduces overall system complexity while maintaining full automation capability, as the network learns all detection and assignment relationships in a single integrated framework.
4Loss of information
If objects are far away or paths have high curvature, then the DNN must interpret limited pixel information, but accuracy of spatial location and size determination decreases
Solution Approach 1:
The unified neural network acts as an intermediary that processes raw sensor data directly without relying on limited pixel interpretations. By operating on the original high-dimensional sensor input rather than derived 2D projections, the network preserves spatial information even for distant objects or curved paths. The model learns to extract accurate spatial relationships directly from the sensor data, bypassing the information loss that occurs in traditional projection-based approaches.
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. Additionally, a machine may perform control operations based at least on the path assignments.