YOLOv7 with Multiple Object Tracker - SORT Algorithm about yolov7_openvino_cpp-python HOT 10 CLOSED

No, I think YOLOv7's backbone is natively not for object tracking. You should combine it with an extra reidentification model to achieve this task.

from yolov7_openvino_cpp-python.

Hopefully this doesn't sound silly but this old pyimagesearch blog post (and code) talks about using a object detection classifier something like once every 30 frames and then object tracking algorithm on all other frames to increase FPS performance. I think the basics is the object detection running only every 30 frames just grabs bounding box coordinates and then passes it off to the tracking algorithm that runs majority of all other frames.

The code in the pyimagesearch repo:

	# check to see if we should run a more computationally expensive
	# object detection method to aid our tracker
	if totalFrames % args["skip_frames"] == 0:
		# set the status and initialize our new set of object trackers
		status = "Detecting"
		trackers = []

		# convert the frame to a blob and pass the blob through the
		# network and obtain the detections
		blob = cv2.dnn.blobFromImage(frame, 0.007843, (W, H), 127.5)
		net.setInput(blob)
		detections = net.forward()

Sort of curious on your thoughts if the Yolo v7 tiny (same thing I have been experimenting around with here) could be the more computationally expensive part that only runs once every 30 frames or so....and then use some sort of tracking algorithm that runs on other frames....? Ultimate goal would be for counting people going in and out crossing a line like the pyimagesearch tutorial....and hopefully enough FPS to accommodate walking people through the frame of a building entry way.

Im sort of curious to replicate the pyimagesearch code that uses dlib for the object tracking and then maybe the Intel tiny yolo v7 as the object detection part.....this whole journey started with the pyimagesearch code where if the model they used on the object detection side was good, like it worked for detecting people from different angles I would have never have had the need to find this repo : )

Just curious on your thoughts from hi level perpective....? Like if dlib is worth using for some concept idea....

• http://dlib.net/
• http://dlib.net/correlation_tracker.py.html

from yolov7_openvino_cpp-python.

Hi @bbartling dlib is a great library tool. I think your idea should work, but you need to be careful on the lost object or new object tracking between two frames.
So far, PP-tracking could be another choice for you, and OpenVINO now can directly load a PaddlePaddle static model:
https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.5/configs/mot/README_en.md

from yolov7_openvino_cpp-python.

Cool I may tinker around with this running object classification on a yolo v7 every 30 frames and some sort of object tracking on the other frames....either dlib or PaddlePaddle like you linked above.

from yolov7_openvino_cpp-python.

Hi @bbartling I create a new notebook for person tracking based on DeepSORT algorithm.

from yolov7_openvino_cpp-python.

AWESOME, ill have to check it out! You run in Windows, MAC, or Linux? Thanks for sharing!!

Does it work like the pyimagesearch blog from a few years back where every 30 frames or so object detection runs (Yolo v7?) and then the vast majority all other frames object detection runs ? At least they state in the blog it improves FPS and processing performance greatly vs running object detection on every video frame...

from yolov7_openvino_cpp-python.

No, it run detection on each frame, however I use a lighter detection model in this example, and its accuracy and performance are good for a middle range object (over 60FPS with 11th gen CPU). If you wish to get a more accurate result on small object, you can switch the model and code to YOLOv7's.

from yolov7_openvino_cpp-python.

@OpenVINO-dev-contest could this ever be developed into processing multiple video streams at once to count people going in and out if the computer vision app can be tapped into the network handling security camera feeds???

For example this DVR managing the IP cameras on all building entrances/exits, 4 in total:

from yolov7_openvino_cpp-python.

@OpenVINO-dev-contest could this ever be developed into processing multiple video streams at once to count people going in and out if the computer vision app can be tapped into the network handling security camera feeds???

For example this DVR managing the IP cameras on all building entrances/exits, 4 in total:

Of course yes, you can put your RTSP address of IP camera directly into OpenCV function to decode a video stream, for example: cap = cv2.VideoCapture('rtsp://192.168.0.1/video.264'), and create multiple workload for your NVR device. The format of RTSP address depends on which IP camera vendor you use.

from yolov7_openvino_cpp-python.

wow cool stuff...this would be a really neat project

from yolov7_openvino_cpp-python.