Real-time detection using YOlOv5

[mana747]

I want to perform real-time detection using Parrot Anafi and YOLO v5. I’m using rtsp, but I get the error “Failed to open rtsp: // …”. The log is shown below. Please let me know if anyone has been able to resolve this error. Thank you.

[Fabien.esigelec]

and after whit threating

from time import sleep
from matplotlib.image import imread
import torch
import cv2
import threading
import os

RTSP_URL =‘rtsp://192.168.42.1/live’
os.environ[‘OPENCV_FFMPEG_CAPTURE_OPTIONS’]=“rtsp_transport;udp”

Model

model = torch.hub.load(‘ultralytics/yolov5’, ‘yolov5s’, pretrained=True)

THICKNESS = 2
COLOR = (255, 0, 0)

make a thread for finding human

class MyThread (threading.Thread):
def init(self):
threading.Thread.init(self)
self.result = None
self.frame = None
self.stop = False

def run(self):
model = torch.hub.load(r’/home/fabien/.cache/torch/hub/ultralytics_yolov5_master’,‘custom’,path=r’/home/fabien/projetS8/drone/yolov5s.pt’,source=‘local’)
while not self.stop:
# wait if no frame
if self.frame is None:
sleep(0.01)
continue

        # Inference
        results = model([self.frame]).xyxy[0]

        # take one people
        for row in results:
            if not row[-1]:
                self.result=tuple(row.int().numpy()[:-2])
                self.frame = None

open camera and check if the webcam is opened correctly

cap = cv2.VideoCapture(RTSP_URL)
if not cap.isOpened():
raise IOError(“Cannot open webcam”)

create thread and start

myThread = MyThread()
myThread.start()

while True:
# get next frame
ret, frame = cap.read()

# send new frame 
if myThread.frame is None:
    myThread.frame = frame

# draw rectangle
if myThread.result is not None:
    frame = cv2.rectangle(frame, myThread.result[:2], myThread.result[2:], COLOR, THICKNESS)

# show the frame
cv2.imshow('Input', frame)

# read escape key 
c = cv2.waitKey(1)
if c == 27:
    myThread.stop=True
    break

close webcam

cap.release()
cv2.destroyAllWindows()

[Fabien.esigelec]

after that I take an image out of 4 and I can integrate my image processing with a counter and I use pcmd movements. I would like to use import Queue but it’s still complicated for me

[mana747]

Thank you for your reply.
The code you gave me worked fine.
However, there is a phenomenon that the image is delayed and roughened.
Therefore, the detection accuracy is low.
Is such a phenomenon occurring in your environment?