Hello Tello Pilot!
Join our DJI Tello community & remove this banner.
Sign up

Object detection using DJI Tello

Joined
Mar 13, 2021
Messages
5
Reaction score
3
I am not using YOLO or any pre-trained model.
Using Tensorflow API with Python to train a model on custom image.
This works very well with my laptop camera but with DJI Tello's camera I get
"circular buffer overrun. to avoid, increase fifo_size url option. to survive in such case, use overrun_nonfatal option" and then tello's camera stops working.
I tried with "?overrun_nonfatal=1&fifo_size=100000000", now it doesn't blow up but doesn't work either.
It just renders one frame then freezes.

My model is too heavy around 5 GB in size.
When I don't try to detect object in the image then streaming works alright but as soon as I call my object detection method, it starts showing above issue.
I am passing image.copy() in detection method, so not sure what is the issue.

Any help? any one?
 

SyedQasm

New member
Joined
Apr 5, 2020
Messages
2
Reaction score
0
Hi ramkinkarpandey,
  1. Is there any reason for not going with the Yolov4 or Maybe Yolov5?
  2. Is that object detection model is giving you more FPS?
  3. Which classes are you detecting with the custom model? Is that just a person?
BTW I am working on this using Yolo v5s. So any comment will be highly appreciated. :)
 
Joined
Mar 13, 2021
Messages
5
Reaction score
3
Hello SyedQasm,
I know that Yolo V4/V5 are faster but doesn't feel like industry standard (it could be just me) they aren't MIT licensed. So many of the companies don't like to use YOLO.
Moreover some says that YOLO V5 is a scam :)
I was able to solve the issue using the pointer given to me by "jefflroberts". Issue my model is taking 0.4 seconds in prediction and in that time I am not reading any images from stream. But Tello is still stream so after couple of seconds streamed frames (which are not read and still in buffer) goes out of buffer thus I get "circular buffer overrun" issue.
For handling this, I started a thread, in main thread I was reading frames and in another thread I was performing prediction. Both used queue to share frames. So removed dependency on processed from reading.
I am trying to control my drone using hand gesture commands, so multiple hand signs (left, right, takeoff, land etc)
Regards,
 

SyedQasm

New member
Joined
Apr 5, 2020
Messages
2
Reaction score
0
Hello ramkinkarpandey,

Firstly I am more concerned about the FPS so that my drone can be able to do quick movements, avoid any close obstacle and all that, On the other hand, you are concerned about the industry standard, that seems totally fine to me, it's just the priorities and Use Case.
Yes, YoloV5 is not from the Previous Yolo Author "pjreddie", Who published the YoloV4 with the paper and considered it as an Official Version using Darknet. YoloV5 is by some "Ultralytics" Guys aiming to provide better speed/with ease of usage/smaller model and so it's FAKE.
So YoloV5 "s" version suits my use case or anyone who wants to get more speed from this little beast.
Anyways I am also training on my custom dataset for multiple gestures-based flight.

Regards,
Syed Qasim
 

jefflroberts

Member
Joined
May 29, 2019
Messages
8
Reaction score
5
I know I answered this initially, and I am glad I could help, but I know very little about computer vision. I would really like to learn more, but I don't know where to start. I have a non-edu tello and I also have a bigger drone with an onboard raspberry pi and I have several pi cameras. Could anyone direct me to a beginner tutorial to get started or some rather simple code?
 
Joined
Mar 13, 2021
Messages
5
Reaction score
3
I know I answered this initially, and I am glad I could help, but I know very little about computer vision. I would really like to learn more, but I don't know where to start. I have a non-edu tello and I also have a bigger drone with an onboard raspberry pi and I have several pi cameras. Could anyone direct me to a beginner tutorial to get started or some rather simple code?
Jeff,
Start with

Lot's of videos with code and explaination.
 
  • Like
Reactions: jefflroberts

New Threads

Members online

Forum statistics

Threads
5,073
Messages
36,693
Members
13,642
Latest member
HillClimber