New Delhi: Autonomous driving is now coming up with vehicles that can operate independently with different levels of autonomy ranging from level 1 with driver assistance to level 5 with complete self-driving under all conditions. The complexity of autonomous driving requires the continuous tuning of different perception models for a variety of road scenarios. To ensure best performance, data plays an essential role in every stage of product development. Big data or petabytes is vital, and it is even more important to have the right data for the right use cases at the right time. Let us look at the significance of data analytics for autonomous driving.Application of data analytics
In autonomous driving, data analytics can be applied at both active and passive levels. Active analytics is when different techniques are used to collect the right data for the right use-cases, and it is more direct and saves cost. Passive analytics uses techniques to find the right data among petabytes of collected data and this incurs more costs and time.
In order to reach level 5 autonomy in ADAS, more software development, simulation and validation, along with AI and high computing are needed. The right data will be the key differentiator here. Developing, validating, and ensuring high performance of the autonomous driving software using AI needs millions and billions of km of real world scenarios that is neither economically nor practically feasible in the automotive ecosystem of today. Data analytics has a huge potential to collect, process and store right data for right use case at right time in an optimal manner. This will save time, resources and money for OEMs and automated driving solution providers.
Data in ADAS
Sensors that are fitted on the vehicle such as Cameras, Radars and LiDAR collects the ADAS test data and captures and store real-time road scenarios while driving. The data that is generated is massive. One km of driving using a standard setup of sensors can generate gigabytes to terabytes of data. That’s not all, millions of km of data (petabytes of raw data) need to be gathered to validate the functionality of the autonomous diving software. This poses some challenges that need to be addressed as solutions must be efficient in a short time. The overall flow of ADAS test data requirements include data collection, data upload, data wrangling, data enrichment & search and data usage.
Some of the challenges are:
Active and passive analytics: End to end solution.
As the degree of autonomy grows, the need for data also rises. There is a lot of data scarcity in newly collected data. Data shortfalls often get revealed while developing the autonomous driving software development and validation. This means loss of time and money. Active analytics help obtain the right data at the right time for the right purpose.
Passive analytics, on the other hand, is applicable where petabytes/millions of km of AD Test data has already been collected, stored, and processed. Each gigabyte of test data is vital and costly to store and process, whether a data centre or in cloud. As a result, it is critical to process each gigabyte of data and categorize it based on its value. If the data has value, it has to be investigated further to assess its precision to precisely locate the road environment and its participants (may be static or dynamic) by fusing data from various sensors. Passive analytics can also help to increase data re-use, and in some circumstances, the cost of gathering new data can be reduced but with certain limitations. The precise data discovered has to be made available and easily accessible for use. This increases the lead time for AD software development and validation.
Data Collection: To collect data, the fleet vehicles are fitted with the required sensors including cameras, radars and LIDARs to capture real-world road scenarios. The dataset used here is real-time data sourced from active analytics. Employing Map database and routing algorithms, the system generates optimal routes based on data requirements. This facilitates the collection of relevant data for specific use cases at the right time and location.
Map and weather services: Map and weather services (passive analytics) play a crucial role in automatically enhancing raw data by providing information on various environmental conditions such as country, road type, tunnels, weather, light conditions, and more. These attributes are then made accessible to end-users for preparing datasets tailored to their specific use cases, including AI training, simulation, validation, and more.
Traffic participants detections: On roads, there are two types of traffic participants: static that includes road signs and dynamic, including cars, trucks and pedestrians. Trained neural networks are employed for the automatic detection of these participants through passive analytics. Subsequently, these attributes are made available to end-users for preparing datasets tailored to their specific use cases, such as AI training, simulation, validation, and more.
Scenario detection: Achieving a higher level of autonomy necessitates the identification of complex road scenarios. Leveraging image captioning and clustering techniques, the overarching scenario can be detected. This facilitates the search, analysis, and preparation of datasets based on these scenarios through passive analytics. The categorized flow of scenarios is employed for seamless accessibility during development and validation processes.
Meta Data Lake: The insights or the meta data gathered from different steps are consolidated and stored within a centralized database known as the “Meta Data Lake.” Subsequently, the Data Search & Analytics Engine consumes this meta data for further analysis and retrieval.
Smart image search: As part of passive analytics, the features of uploaded data images are extracted and stored in a dense vector database. When a user inputs or uploads an image into the search engine, the features of the uploaded image are extracted on the fly. These features are then compared within the dense vector space, returning the nearest neighbor images from the vector database as results along with a relevance score.
Overcoming ADAS test data challenges
Ensuring AD/ADAS solutions are efficient, best-in-class, and reach the market quickly depends on conquering three key test data challenges.
1.Data collection: Gathering the right data, at the right time, for the specific use case is crucial. This involves capturing relevant information like the road environment, static and dynamic traffic participants, with high accuracy and completeness.
2.Data processing: Transforming raw data into actionable insights is vital for AD software development and validation. We need to extract precise information, such as road features and the behavior of other vehicles, from the collected data stream.
3.Data search: Making raw data insights readily accessible is essential for efficient data selection and quick retrieval of the right information. This facilitates faster analysis and decision-making during the development process.
Results based on different searches
Text-based search: Following passive analytics processes, the text-based search manifests in the search engine, utilizing Map database for filtering based on country, road type, and environmental conditions.
Scenario-based search: The search engine reflects scenario-based search outcomes after passive analytics processes, incorporating image captioning and clustering algorithm techniques. This approach expedites the preparation of datasets, offering efficient results within a short timeframe.
Image-based search: Image-based search within the search engine is an integral aspect of passive analytics. It involves searching and analyzing the visual content of input images against a vector database, which is constructed as part of the passive analytics process.
Conclusion
In the realm of autonomous driving, the adept utilization of Data Analytics is the key. The intricacy of Autonomous Driving (AD) and Advanced Driver Assistance Systems (ADAS) software development, simulation, and validation processes escalates proportionally with the level of autonomy. The meticulous application of both active and passive data analytics proves instrumental in addressing the myriad challenges associated with the development and validation of AD/ADAS products. By harnessing the power of data analytics, the automotive industry can not only pinpoint the right data but also ensure that it aligns seamlessly with the evolving demands of autonomous and connected driving technologies, paving the way for a safer and more efficient future on the roads.
(Disclaimer: Surendhar Selvaraj and Jaydeep Banerjee, Autonomous Mobility, Continental Automotive India. Views are personal.)
