Standardized annotated dataset generation from raw data collected by lidar sensor

Advanced Driving Assistance Systems and Autonomous Vehicles could significantly reduce road accidents that can be attributed to the driver in more than the 90% of the cases. One of the biggest challenges in driving automation is to ensure the system is safe before placing it on the road since they can perform badly in certain situations such as system fault, inclement weather conditions and complex driving environment, but also cyber security comes to the attention when vehicles are automated. The solution to achieve safety in the context of ADAS/AVs is testing the systems to validate them, in particular driving tests are needed. It has been estimated that vehicles must be driven for hundreds of million kilometers in order to achieve the expected result for the certification process and it would take tens of years. For this reason, virtual testing is the preferred solution by the industry. Virtual testing reduces costs, improves repeatability and raises the scale of the number of tests. The data-driven approach gives a medium fidelity representation of the real world but high quantity of processed data is needed. An important process is the data annotation that is mainly done by hand for its reliability but is slower and more expensive than an automated annotation. As a consequence, collecting and processing a large amount of data needed for the validation is a costly process, both in terms of time and money, therefore the trend is to concentrate on the critical scenarios, but also to standardize databases in order to increase the exchange of data across company and organization and in addition to automate processes. Standardization of databases means data are available for more users, accelerating the automotive development, testing and validation. The aim of this thesis study is to report the process of developing a tool capable of automatically generating standardized labels for an automotive raw dataset. The research and developing processes have been done as part of an internship in Concept Quality Reply. The developed tool is able to take as an input a structured dataset, collected from the real world by a standardized sensor setup. The dataset’s structure is compliant to the one developed for NuScenes by Motional which is a multimodal large-scale dataset for autonomous driving that is now leading the industry. In particular the data under analysis are point clouds generated from a lidar sensor. The developed system recognizes pedestrians and vehicles with a certain accuracy and generating bounding boxes and labels. The 3D object detection process is made through a convolutional neural network model that gives as an output the labels with their confidence level. The pipeline ends with a conversion process of the resulting labels into a standardized format. The standard identified is OpenLabel, defined by ASAM (Association for Standardization of Automation and Measuring Systems) that is the current de-facto standard in the automotive industry. To permit the visualization of the labels, a graphical user interface has been developed together with a web application and an APIs collection, that make possible the communication of the user with the tool hosted by a server and shapes the developed system as a “Software as a Service” (SaaS).