Development and integration of OEDR module with synthetic dataset generation for self-driving cars

Abstract

The fundamental motivation for the development of self-driving cars is safety, yet upholding the highest safety standards is challenging due to the complex and dynamic nature of real-world driving environments. For autonomous cars to operate reliably, they must be tested over millions of kilometers, necessitating a substantial financial investment, logistical planning, and infrastructure development. Furthermore, real-world testing is limited by its inability to replicate the full spectrum of edge cases and rare event scenarios that autonomous cars may encounter. The creation of var-ied and controlled driving scenarios necessary for thorough testing is made possible by synthetic datasets. These datasets enable the creation of diverse, repeatable, and parameterized driving scenarios, which are critical for validating the performance and robustness of autonomous systems under varying conditions. The majority of the acci-dents are due to the incorrect or delayed perception of the environment. This study fo-cuses on the development of an Object and Event Detection and Recognition (OEDR) module integrated into an ego vehicle in a customized simulator environment where a wide array of critical traffic situations, including occlusions, varying illumination, dynamic obstacles, and different combinations of weather conditions, are replicated. With the usage of state-of-the-art deep learning models, the OEDR module guaran-tees precise object detection and recognition in diverse crucial road conditions in real time. The annotated synthetic dataset generated through this simulation framework seeks to improve the reliability of autonomous cars.