Overview
Scene Intelligence with Rosbag on AWS is purpose-built to help streamline the development process for Advanced Driver Assistance Systems (ADAS) and Autonomous Vehicles (AV). The solution features modules for sensor extraction and object detection, helping machine learning engineers and data scientists to accelerate scene search for model training.
You can use this solution to stage sample rosbag files, extract rosbag sensor data such as metadata and images, apply object detection and lane detection models to the extracted images, as well as apply and store scene detection business logic.
Benefits
Scalable, flexible data pipelines that reliably ingest, transform, label, and catalog billions of miles of real or simulated data.
Greater accessibility for global teams to search, identify, and analyze automotive data.
Reduce the number of dependencies and prerequisites with open-source configuration options.
Technical details
You can automatically deploy this architecture using the implementation guide and the accompanying AWS CloudFormation template.
Step 1
The AV uploads the rosbag file to HAQM Simple Storage Service (HAQM S3). The end user invokes the workflow to start processing through HAQM Managed Workflows for Apache Airflow (HAQM MWAA) and a directed acyclic graph (DAG).
Step 2
AWS Batch pulls the rosbag file from HAQM S3, parses and extracts the sensor and image data, and writes this data to another S3 bucket.
Step 3
HAQM SageMaker applies object detection and lane detection models to the extracted data. SageMaker then writes the data and labels to another S3 bucket.
Step 4
HAQM EMR Serverless (with an Apache Spark job) applies business logic to the data and labels in HAQM S3. This generates metadata related to the object detection and lane detection. HAQM EMR Serverless then writes the metadata to HAQM DynamoDB and another S3 bucket.
Step 5
An AWS Lambda function publishes new incoming DynamoDB data (metadata) to the HAQM OpenSearch Service cluster. The end user accesses the OpenSearch Service cluster, through a proxy on HAQM Elastic Compute Cloud (HAQM EC2), to submit queries against the metadata.
Step 1
The AV uploads the rosbag file to HAQM Simple Storage Service (HAQM S3). The end user invokes the workflow to start processing through HAQM Managed Workflows for Apache Airflow (HAQM MWAA) and a directed acyclic graph (DAG).
Step 2
AWS Batch pulls the rosbag file from HAQM S3, parses and extracts the sensor and image data, and writes this data to another S3 bucket.
Step 3
HAQM SageMaker applies object detection and lane detection models to the extracted data. SageMaker then writes the data and labels to another S3 bucket.
Step 4
HAQM EMR Serverless (with an Apache Spark job) applies business logic to the data and labels in HAQM S3. This generates metadata related to the object detection and lane detection. HAQM EMR Serverless then writes the metadata to HAQM DynamoDB and another S3 bucket.
Step 5
An AWS Lambda function publishes new incoming DynamoDB data (metadata) to the HAQM OpenSearch Service cluster. The end user accesses the OpenSearch Service cluster, through a proxy on HAQM Elastic Compute Cloud (HAQM EC2), to submit queries against the metadata.
Related content
This Guidance demonstrates how customers can process and search high-accuracy, scenario-based data with the Autonomous Driving Data Framework (ADDF).
- Publish Date