Synthetic data refers to computer-generated data designed to mimic real-world data, often with a focus on specific scenarios or use cases. It is becoming increasingly important for a wide range of applications. From training self-driving car algorithms to testing safety features in new vehicle models, it is a powerful tool for the automotive industry and is helping to accelerate development.
Applications of synthetic data
Synthetic data has proven to be an invaluable tool in the development of self-driving cars, which rely on sensors and machine learning algorithms to operate safely and accurately.
Companies like Tesla and Waymo have access to vast amounts of real-world data required to train these algorithms, but many other companies struggle to obtain sufficient data. It offers a practical solution by allowing developers to create an almost unlimited amount of synthetic data that can be used to train these algorithms. This enables developers to test and refine their algorithms without relying solely on the limited amount of real-world data available to them.
The automotive industry was one of the first proponents of synthetic data, but its focus has mainly been on autonomous driving aspects. However, the focus now should expand to other areas, such as in-cabin monitoring and around-car monitoring for security and safety, which is vital as we make the transition from non/semi-autonomous to fully autonomous vehicles. In addition, synthetic data can be used to train algorithms that recognize license plates and street signs.
Synthetic data can be used to simulate crash scenarios, empowering engineers to test the effectiveness of safety features without the need for real-world testing. This means that engineers can also test the effectiveness of safety features in a more comprehensive way.
Here are a few aspects of development that must be taking into consideration:
- Simulation and virtual testing: Automotive engineers use synthetic data to create virtual environments that simulate real-world driving conditions. Synthetic data helps generate diverse scenarios, including different weather conditions, traffic situations and road configurations;
- Training machine learning models: Synthetic data is used to train machine learning algorithms and neural networks. It helps augment limited real-world data sets by generating additional training samples. This augmentation improves the accuracy and generalization of models, making them more robust in different situations;
- Anomaly detection and fault diagnosis: Synthetic data aids in testing and validating anomaly detection systems in vehicles. By generating diverse synthetic data representing various fault conditions, engineers can train algorithms to identify and diagnose anomalies in the vehicle’s systems;
- Sensor development and calibration: Synthetic data assists in developing and calibrating sensors used in autonomous vehicles, such as lidar, radar and cameras. Generating synthetic sensor data enables engineers to analyze and fine-tune sensor parameters.
Current bottlenecks
There are numerous challenges associated with utilizing synthetic data. Generating synthetic data that accurately reflects real-world driving conditions, including variations in lighting, weather and object interactions, remains a challenge. Ensuring models trained on synthetic data perform well in real-world scenarios requires thorough validation and generalization testing. Adhering to privacy regulations while maintaining data diversity and realism in synthetic scenarios presents a balancing challenge. Accurately representing sensor characteristics and variability, such as lidar and radar behavior, in synthetic data is crucial for reliable testing and calibration. Generating large-scale synthetic data sets and running complex simulations demands significant computing resources and efficient processing.
Adequately representing infrequent exceptional driving situations in synthetic data is challenging but necessary for comprehensive testing. Developing high-quality synthetic data sets and simulation environments can still be resource-intensive, requiring careful cost-benefit considerations.
Next big developments
We anticipate multiple developments at Mindtech. Breakthroughs in deep learning, reinforcement learning, natural language processing and computer vision, leading to broader applications across industries. A focus on robust technologies like encryption, secure computing and secure data sharing to address data privacy and security concerns. Advancements in sensor tech, network infrastructure and data processing to support the proliferation of IoT devices and seamless connectivity. We anticipate developments in generative AI models that can create realistic and novel content, such as images, videos and text, with wide-ranging implications for creative industries and content generation.
Mindtech’s approach
Mindtech has taken an ‘MLops’ type approach to the creation of synthetic data, that is to say, we look at the whole lifecycle. We start the process by an analysis of current data using our data analysis platform, which analyzes for appearance, content and neural network fit. This enables the user to rapidly identify gaps and bias within the data sets and guide the next stage of our MLops flow, which is the specification of the scenes and assets required to “fill the gap”. Mindtech has created a UI tool for easy creation of the scenarios that will create the data. This is important as the nature of MLops is an iterative process that improves iteration upon iteration. From here we simulate, using a behavioural-led simulator, that can automate the creation of large amounts of relevant data.
Within the platform we include the ability to generate both ‘invariant’ and context data. Invariant data is where we randomize the background, with specific items of interest rotated and relit to get a thorough understanding of that object (object could be a person, item, vehicle, sign etc.). Context data is where we simulate the real-world context of where the system will be deployed, for example a car driving down a city street.
A key point is that for all our data we perform domain matching. That is to say, we match the characteristics of the deployment environment. So for example, we perform colour matching of the environment, match lens distortions and other camera pipeline characteristics. This ensures we get the most accurate results for training.
Synthetic data is an increasingly important tool in the automotive industry, enabling researchers and engineers to accelerate development, improve safety and test new technologies more effectively. By embracing this technology, the automotive industry can continue to innovate and evolve, paving the way for a safer and more sustainable future.
