Developer Documents Real-Time CARLA Perception Stack Using YOLOv11, ROS 2 and ByteTrack

A developer spent three weeks building a real-time robotic perception system inside the CARLA driving simulator, integrating YOLOv11 for object detection, ByteTrack and StrongSORT for multi-object tracking, and ROS 2 as the middleware layer. The project aimed to detect and track objects across frames with stable identities while feeding structured data into a ROS 2 navigation stack fast enough for autonomous decision-making. A missing ROS 2 clock timestamp in the camera callback caused downstream timing mismatches that made the tracker behave erratically, a bug that took two days to diagnose. A second key issue arose because blocking inference calls on the ROS 2 subscription thread caused stale frames to queue up, which was resolved by decoupling ingestion from inference using a bounded queue and a dedicated inference thread. The write-up focuses on practical failures and fixes rather than clean architecture, offering lessons for engineers building real-world robotics perception pipelines.