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

A new runtime controller called RAMS dynamically selects among three YOLOv8 object-detection tiers on embedded hardware, aiming to balance inference latency and detection quality under fluctuating resource pressure without model-reload delays [1]. The lightweight controller monitors device pressure and calibrates switching thresholds from idle behavior, choosing between NANO, SMALL, and MEDIUM YOLOv8 variants at 320, 416, and 640 pixels, respectively [1]. RAMS defines five switching policies, including two detection-conditioned variants that prevent aggressive downgrades after recent vulnerable-road-user (VRU) detections [1]. The authors also introduce the VRU-Weighted Accuracy Score, or SWAS, a scalar metric designed for offline policy comparison that does not require ground-truth annotations [1]. An oracle-bounded variant of SWAS separates detector circularity from genuine tier-retention benefit [1]. Across Raspberry Pi 5, x86 laptops, and Jetson Orin ONNX and TensorRT deployments, the same controller equations operate over a 37x latency range [1]. On a Jetson Orin running TensorRT under heavy load, the safety2 policy achieves a mean latency of 3.41 ms, which is 5.6 times faster than fixed-MEDIUM inference, while retaining 74 percent of its proxy accuracy through near-NANO operation with selective SMALL and MEDIUM locks during VRU-positive windows [1]. Detection-conditioned switching improves SWAS by 25.4 percent under oracle scoring and by 47.3 percent under detector-derived scoring relative to threshold-only policies under heavy load [1]. Live evaluation on the KITTI dataset reports per-tier VRU recall of 24.2 percent, 41.2 percent, and 59.0 percent, showing that reactive overrides are fundamentally limited by baseline detector recall [1]. The work appears on arXiv in the computer-vision and pattern-recognition category, where papers can be accompanied by interactive demos through an integration with Hugging Face Spaces that lets readers try models directly in a browser [4][5].