AI for Environment Perception in Automotive Domain. A Necessity for Automated Vehicles

With the rise of new forms of mobility and the shift towards automated driving systems, environmental perception is becoming a key issue in ensuring safe and reliable travel. Advanced driver assistance systems (ADAS) and automation technologies rely on increasingly sophisticated sensors, whether they are installed in vehicles or integrated into road infrastructure. The effectiveness of these perception systems depends directly on the ability of sensor data processing algorithms to accurately, reliably, and robustly estimate the state of the five key elements of the road scene: obstacles, infrastructure, ego-vehicle, environment, and driver. The fusion of data from exteroceptive and proprioceptive sensors makes it possible to build local dynamic perception maps (CPDL), which are essential for activating critical features such as lane tracking, trajectory change assistance, emergency braking, safety distance regulation, and automated intersection and parking maneuver management.

In this context, ensuring the performance and quality of perception systems becomes a top priority. This involves continuous optimization of vision technologies, as well as fusion and machine learning algorithms to ensure reliable operation regardless of environmental conditions. Only high-precision perception, using both onboard data and data from connected infrastructure, will accelerate the adoption of automated vehicles and significantly improve user safety.

This article offers an in-depth exploration of the various methods and approaches used to process data from onboard sensors. It covers obstacle detection techniques, moving object tracking, traffic sign recognition, and algorithms for assessing risks and anticipating dangerous road conditions. For example, early detection of adverse weather conditions, such as rain or snow, plays a key role in adapting automated driving strategies to ensure passenger safety. Similarly, recognition of traffic signs and road markings is essential to the proper functioning of automated driving systems, which must interpret and react instantly to this information.

In addition, this article examines the impact of advances in image processing, with a particular focus on the rise of deep learning algorithms, which have significantly improved the accuracy and efficiency of embedded vision systems. Convolutional neural networks (CNNs), for example, have enhanced the ability of vehicles to accurately identify and classify complex objects on the road, with performance continuously improving as models are refined. Deep learning also allows for greater adaptability, as models are trained on countless scenarios, making systems more robust in varied and dynamic environments.

Through an in-depth review of scientific...