A Motion Gated Intrusion Detection Pipeline Using NVIDIA DeepStream on Jetson

Edge AI video analytics is an increasingly active research and deployment area, driven by growing demand for privacy-preserving, locally operated surveillance solutions with advanced alerting capabilities. Deploying such systems on embedded hardware introduces non-trivial challenges: cost and power constraints, the need to sustain multiple concurrent streams, and the computational demands of modern detection models must all be addressed locally. This thesis presents a real-time multi-stream video intrusion detection pipeline deployed entirely on the NVIDIA Jetson Orin Nano Super via the DeepStream SDK, designed to fully exploit the platform’s dedicated hardware accelerators. A motion-based pre-filtering stage is proposed based on the MOG2 background subtractor, implemented as a native zero-copy GStreamer plugin via NVIDIA VPI, which selectively suppresses frames carrying no relevant motion before they reach the inference engine. Compared with a baseline Python implementation, the DeepStream pipeline achieves speedups of up to 4× without MOG2 and 3.3× with it. This work characterizes YOLOv8 and YOLOv11 across four model sizes, four to ten parallel streams, and three precision modes (FP32, FP16, INT8) on the VIRAT dataset. Evaluation covers throughput, mAP@50, and Time To First Detection, a metric introduced in this work to quantify the latency in frames between an object’s appearance and its first valid detection. Results show that INT8 yields a significant speedup at the cost of severe accuracy degradation. FP16, leveraging the Tensor Cores of the Jetson platform, proves simultaneously faster and more accurate than FP32 and is the clear deployment choice. The gated configuration outperforms its ungated variant below 50% motion density and sustains eight parallel streams with YOLO11-M FP16 well above the IEC 62676-4 minimum requirement of 12.5 fps/stream for video surveillance systems, demonstrating that a fully local, privacy-preserving deployment is feasible.