Browsing by Author "MEHAYA, Mohammed Elmehdi"

Now showing 1 - 2 of 2

Sheep Detection, Tracking and counting from aerial images using Deep Learning
(جامعة غرداية, 2021) MEHAYA, Mohammed Elmehdi; DJEKABA, Fatima
Object detection is widely used in the field of computer vision. Furthermore, it can be harnessed in agriculture and farming, especially with the new methods that achieve promising results. Nowadays, the problem is tackled using either traditional machine learning methods that use computer vision techniques or deep learning methods. In this work, we investigate the deep learning stateof-the-art tools to create a smart system for detecting, tracking and counting sheep using aerial images captured by a drone. In the process, we gather sufficient data with good quality and use it to train a model dependent on the YOLOv4 network. Next, we tackle the counting stage directly using an innovative method that uses an imaginary line cutting the processed frame incrementing the counter whenever an intersection between the bounding box and the gate happens. However, we had to introduce an intermediate stage because of low performance. That intermediary is called tracking. The results obtained by the experiment are highly promising in detection with an mAP of 71% and 16.1274 % of avg loss function.
Sheep Detection, Tracking and counting from aerial images using Deep Learning
(université Ghardaia, 2021) MEHAYA, Mohammed Elmehdi; DJEKABA, Fatima
Object detection is widely used in the field of computer vision. Furthermore, it can be harnessed in agriculture and farming, especially with the new methods that achieve promising results. Nowadays, the problem is tackled using either traditional machine learning methods that use computer vision techniques or deep learning methods. In this work, we investigate the deep learning stateof-the-art tools to create a smart system for detecting, tracking and counting sheep using aerial images captured by a drone. In the process, we gather sufficient data with good quality and use it to train a model dependent on the YOLOv4 network. Next, we tackle the counting stage directly using an innovative method that uses an imaginary line cutting the processed frame incrementing the counter whenever an intersection between the bounding box and the gate happens. However, we had to introduce an intermediate stage because of low performance. That intermediary is called tracking. The results obtained by the experiment are highly promising in detection with an mAP of 71% and 16.1274 % of avg loss function.