Augmenting Dynamic Deadline-Driven Model with Deep Learning for Safer Autonomous Driving

Autonomous vehicles (AVs) present significant potential and challenges within transportation systems, necessitating sophisticated computational processes for perceiving environments, decision-making, and control. These processes, organized into AV pipelines, must operate under strict, dynamically varying deadlines to ensure safety and efficiency. Traditional software systems, such as those based on the Robot Operating System (ROS), struggle to manage these dynamic deadlines, often leading to either compromised performance or under-utilization of resources. The Dynamic Deadline-Driven (D3) Execution Model was developed to address these issues by centralizing deadline management and enabling AV systems to adapt computation in real-time based on environmental feedback. However, the D3 model lacks mechanisms for accurately predicting deadlines across diverse scenarios and handling extremely short response times required in emergencies. To address these issues, this paper integrates deep learning models within the D3 framework, enhancing its ability to set precise deadlines and respond effectively in critical situations. By deploying two deep neural network (DNN) models across different modules of the AV pipeline, the augmented system significantly improves safety, predictive accuracy, and latency minimization. Experimental analysis in the CARLA simulation environment demonstrates that the deep learning-enhanced D3 model outperforms traditional approaches, reducing collision rates and improving system responsiveness in long-distance driving and complex urban scenarios.