Computing architecture of an autonomous vehicle
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computing architectures in autonomous vehicles (AVs) face inefficiencies due to the loss of information when transforming radar data tensors into three-dimensional point clouds, leading to reduced resolution and insufficient bandwidth for data transmission to centralized processing devices, and increased load on these devices.
Innovation Solution
Implementing a local data processing device in the radar system to execute a first deep neural network (DNN) that transforms radar data from polar to Cartesian coordinates and extracts embedded features, reducing data transmission to a centralized processing device, which then executes a second DNN for object detection and labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data tensors are transformed into three-dimensional point clouds for centralized processing, then the data can be processed by existing deep neural networks, but information is lost and resolution is substantially reduced
Solution Approach 1:
The radar system performs preliminary feature extraction and generates embeddings from data tensors before transmission. This preliminary processing extracts essential features while discarding redundant information, allowing the centralized system to work with compressed representations rather than raw data, thus preventing information loss during transmission while maintaining detection accuracy
Solution Approach 2:
Embeddings serve as an intermediary representation between raw data tensors and final object detection results. The local radar system transforms data tensors into embeddings that capture essential features, which are then transmitted to the centralized system for further processing into three-dimensional point clouds and object labels, preserving critical information while reducing data volume
2Loss of information
If raw data tensors are transmitted to the centralized processing device, then complete information is available for processing, but bandwidth is insufficient to support the data volume
Solution Approach 1:
The system extracts only the essential features from raw data tensors by generating embeddings at the local radar system. This extraction process identifies and retains critical information while discarding redundant data, enabling transmission of a compressed representation that maintains detection accuracy while dramatically reducing bandwidth requirements
Solution Approach 2:
The system transforms data from one parameter representation (raw data tensors with full dimensionality) to another (embeddings with reduced dimensionality). This parameter transformation maintains the essential information needed for accurate object detection while reducing the data volume to a level that can be transmitted over available bandwidth
3Ease of operation
If three-dimensional point clouds are generated from data tensors, then object detection can be performed, but the resolution is substantially less than lidar-generated point clouds
Solution Approach 1:
The system performs preliminary feature extraction by generating embeddings that capture essential spatial and object information before creating three-dimensional point clouds. This preliminary processing preserves critical measurement information that would otherwise be lost in direct tensor-to-point-cloud transformation, maintaining higher resolution in the final point clouds
Solution Approach 2:
Embeddings act as an intermediary that preserves high-resolution information from data tensors during the transformation to three-dimensional point clouds. The embeddings maintain detailed spatial relationships and object features, enabling the generation of higher resolution point clouds compared to direct transformation methods
4Device complexity
If all processing is performed by the centralized data processing device, then processing is simplified, but the workload and computational requirements are excessive
Solution Approach 1:
The processing architecture is segmented into two parts: local processing at the radar system (generating embeddings from data tensors) and centralized processing (generating three-dimensional point clouds and object labels from embeddings). This segmentation distributes computational workload, reducing the burden on the centralized device while maintaining overall system functionality
Solution Approach 2:
The local radar system performs preliminary processing by generating embeddings from raw data tensors before transmission to the centralized system. This preliminary action filters and compresses the data, reducing the workload for the centralized device while preserving essential information needed for final object detection and labeling
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The technologies described herein relate to a computing architecture in an autonomous vehicle (AV) (100). The AV (100) includes a radar system (102) that includes a local data processing device (204). The AV (100) also includes a centralized data processing device (106). The local data processing device (204) is configured to perform processing on raw sensor data generated by the radar system (102) to form a feature vector. The local data processing device (204) transmits the feature vector to the centralized data processing device (106), which performs further processing to identify an object in an environment of the AV (100).