Computational Performance of ONNX Runtime and NCNN Frameworks for Real-Time Motorcycle Detection on Raspberry Pi 5: A Yogyakarta CCTV Case Study

Abstract

The acceleration of Smart City initiatives across Indonesia has intensified the demand for localized Intelligent Transportation Systems. Although micro-architectures like YOLOv26n offer optimized layer configurations, executing these models for real-time motorcycle detection on standalone edge computers often presents hardware bottlenecks due to software runtime inefficiencies and core thermal stress. This study presents a comparative computational performance benchmarking of the out-of-the-box YOLOv26n architecture across two prominent runtime frameworks: ONNX Runtime and the NCNN framework. The evaluation scope is strictly bounded to hardware resource utilization and physical stability bounds on a Raspberry Pi 5 platform. Telemetry data were logged frame-by-frame at a granular millisecond-level sampling rate over an uninterrupted 194-second capture window under a live HTTP Live Streaming (HLS) traffic feed from the Pasar Gading Traffic Light areas in Yogyakarta. The empirical results demonstrate that while ONNX Runtime achieves a throughput of 4.92 FPS, it subjects the hardware to severe processor strain, locking the mean CPU usage at 95.83% and causing core temperatures to reside consistently within the active thermal throttling zone (80℃ to 85 ℃). Conversely, the zero-dependency NCNN framework successfully alleviates hardware stress by incorporating low-level hand-written ARM NEON SIMD assembly instructions. NCNN accelerates the processing throughput to 7.36 FPS while simultaneously relaxing the average CPU utilization down to 77.61% and lowering the steady-state memory footprint down to a mean of 1,500.13 MB. Ultimately, this benchmark establishes that assembly-level compiler optimizations via NCNN offer the most viable engineering deployment roadmap for standalone edge-based urban traffic surveillance nodes.