Slow Camera with YoloV5

Question

I’m working with a Jetson Orin Nano and a Raspberry Pi Camera Module V2, aiming to run YOLOv5 for real-time image processing. I’ve tried using the following code, but the camera feed is extremely slow and laggy, which I didn’t expect from this setup. How to improve this?

Here is my code:

import cv2
import torch
import numpy as np  

# Load YOLOv5 model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s')

# GStreamer pipeline for CSI camera
def gstreamer_pipeline(
    capture_width=1920, 
    capture_height=1080,
    display_width=960,
    display_height=540,
    framerate=30,
    flip_method=2,
):
    return (
        "nvarguscamerasrc ! "
        "video/x-raw(memory:NVMM), "
        "width=(int)%d, height=(int)%d, framerate=(fraction)%d/1 ! "
        "nvvidconv flip-method=%d ! "
        "video/x-raw, width=(int)%d, height=(int)%d, format=(string)BGRx ! "
        "videoconvert ! "
        "video/x-raw, format=(string)BGR ! appsink drop=True"
        % (
            capture_width,
            capture_height,
            framerate,
            flip_method,
            display_width,
            display_height,
        )
    )

# Live video capture
cap = cv2.VideoCapture(gstreamer_pipeline(), cv2.CAP_GSTREAMER)

if not cap.isOpened():
    print("Unable to open camera. Please check the connection.")
    exit()

while cap.isOpened():
    cv2.namedWindow("Detect", cv2.WINDOW_AUTOSIZE)
    
    ret, frame = cap.read()
    
    if not ret:
        print("Unable to retrieve frame.")
        break

    # Make predictions using YOLOv5
    results = model(frame)
    cv2.imshow('YOLO', np.squeeze(results.render())) 

    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

Interestingly, when I use the following pipeline, the camera runs much faster:

import cv2
gst_str = "nvarguscamerasrc sensor-id=0 ! video/x-raw(memory:NVMM), width=1920, height=1080, framerate=30/1 ! nvvidconv flip-method=2 ! video/x-raw, format=BGRx ! videoconvert ! video/x-raw, format=BGR ! appsink"

cap = cv2.VideoCapture(gst_str, cv2.CAP_GSTREAMER)

if not cap.isOpened():
    print("Unable to open camera")
else:
    print("Camera opened")

while True:
    ret, frame = cap.read()
    if not ret:
        print("Unable to retrieve frame")
        break

    cv2.imshow("CSI Camera", frame)

    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

I want to use YOLO for real-time image processing and eventually build a training algorithm. I’m open to suggestions for optimizing the current setup or switching to a different YOLO variant or any other approach suitable for the Jetson Orin Nano.