100m Person Detection & Tracking at 35+ FPS on Jetson Orin Nano (8GB)

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

I’m working on a project where I need to detect and track persons up to 100 meters using YOLO v11 on a Jetson Orin Nano (8GB) while maintaining 35+ FPS. Currently, I achieve 16-17 FPS during detection and tracking, but when I move the camera to track a person, the IDs change. To handle this, I implemented ByteTrack for tracking and feature extraction using DeepFace with cosine similarity for re-identification. However, re-identification on CUDA drops the FPS to 1-2.

I have a few questions:

  1. Is YOLO v11 sufficient for detecting persons at 100 meters, or do I need to train a custom model for this range?
  2. If training is required, how should I approach it to ensure high accuracy and efficiency on Jetson Orin Nano?
  3. Which tracking method (ByteTrack vs. OC-SORT) is more reliable for maintaining stable person IDs at long distances?
  4. What optimizations can I apply to maximize FPS without losing too much accuracy?
  5. If this setup is not feasible on Jetson Orin Nano (8GB), which Jetson device would best meet my FPS and range requirements?

I’d appreciate any expert guidance on Jetson AI, real-time tracking, and YOLO optimization.

2

Are you using TensorRT and DeepStream?

You should try TensorRT with int8 and DeepStream if you want better FPS.