• vehicular application
• embedded systems
• data processing
• computer vision
• artificial neural network
• convolutiional neural network
• CNN
• ANN
• artificial intelligence

Artificial Intelligence (AI) and deep learning are gaining in importance due to their potential for a wide range of scientific and industrial applications. Developing AI applications is a complex task with many challenges related to data collection, model training and implementation. As the deep learning become more popular, we witnessed an increased number of intelligent embedded systems such as autonomous vehicles, robots, drones and a plethora of mobile, portable and wearable device that are featured with Artificial Intelligence. However, the performance of DNNs (deep neural networks) running on an embedded system is significantly limited by the scarcity of resources offered by the platform: low computation capability, insufficient memory, limited energy and responsivity are the constraints to be taken into account when designing embedded systems. Their purpose is generally limited to simplistic inference tasks. In this thesis, on-device deep learning algorithms and custom hardware designs are proposed, enabling embedded systems to efficiently perform deep intelligent tasks that are memory-intensive, computationally intensive, and energy-hungry beyond their limited computational resources. This new area of research is known as Embedded Artificial Intelligence. In this thesis, we evaluate architectures, models and implementation problems related to the use of deep learning techniques in the automotive domain. In particular, we focus on different aspects as security, cybersecurity, and maintenance by proposing three case studies to solve these problems. We have developed several deep learning models that help to improve the quality and efficiency of these processes. First, we will focus on the development of embedded vision algorithms for the detection of smoke and fire from camera sensors applied mainly on board train carriages. Vehicular cybersecurity issues are addressed by proposing the development of an Intrusion Detection System (IDS) that can detect in real time cyber attacks on vehicular CAN networks. In the third research study proposes a system able to evaluate the pavement quality of road infrastructure the a real time road classification based on an AI integrated solution. For each vehicular application we provide an analysis of the architecture, datasets and models used, and provide performance metrics for each embedded platform employed referring to real-time performance, memory footprint, power consumption and thermal aspects.