RAMS: Resource-Adaptive and Detection-Conditioned Model Switching for Embedded Edge Perception

The article presents RAMS, a lightweight runtime controller designed for edge object detection on embedded hardware, which dynamically selects among three tiers of the YOLOv8 model (NANO, SMALL, MEDIUM) based on real-time resource conditions without incurring model-reload latency. RAMS implements five switching policies, including detection-conditioned variants that enhance performance during vulnerable-road-user (VRU) detections, achieving a mean latency of 3.41 ms on Jetson Orin TensorRT while maintaining 74% proxy accuracy. This approach is significant for practitioners as it optimizes inference latency and detection quality under variable resource constraints, improving overall system efficiency in embedded AI applications.