ABSTRACT Reliable and time‐bounded perception systems are necessary for autonomous vehicles' (AVs') safe navigation, especially at intersections. In this work, a post‐training quantized you only look once (YOLOv8n) model for real‐time obstacle and traffic light recognition is implemented on the NVIDIA Jetson Orin Nano. The system, which is integrated with the robot operating system 2 (ROS 2) framework, analyzes stereo input from a ZED 2i camera using soft real‐time scheduling theory, taking worst‐case execution time (WCET), jitter, and slack into account. Despite the workload surpassing the traditional schedulability constraints under earliest deadline first (EDF) and rate monotonic scheduling (RMS), empirical evaluation across 423 daytime urban frames reveals that 98.11% of inferences fulfil a 150 ms soft deadline. The results show minimal thermal drift, low jitter, consistent slack margins, and bounded deadline violations (<30 ms). Further analysis incorporates fixed‐priority scheduling, CPU core affinity, and a deadline penalty model to assess safety implications in AV decision loops. While extreme conditions such as night driving or adverse weather were not included, future work will extend to these scenarios. Overall, the findings validate the feasibility of deploying a probabilistically schedulable, timing‐aware perception pipeline for integration into intelligent transport systems (ITS) and edge AI platforms.
M. et al. (Thu,) studied this question.