Wounded Animal Detection in Wild

Story

Large National Parks are always at risk of poaching, sick, and injured animals. Due to the lack of the number of security personnel and technology to survey the enormous national parks, many animals die in the wild untreated. In this project, I tried to develop a real-time wounded animal detection camera system to detect and monitor injured animals in the wild. AMD_xilinx and Canonical delivered us with Ubuntu for Kria KV260, making the job tenfold easier.

Setup Your Kria KV260 Vision AI starter Kit

I have used the link here for setting up the development board.

Setup Edge Impulse on Your Kria KV260 Vision AI starter kit

The link here was used to set up and connect the development board to Edge Impulse.

Then from my host laptop connected to the same network as the Kria, Edge Impulse Studio was opened in a browser where the Kria appeared in the Devices tab. The image data is collected from Google images (.jpg). The photos are of two types:

• Animal with a visible wound (Lable: Flesh wound, No. of images:59)
• Background (Label: Background, No. of image:33)

The images were resized (160x160) using the Impulse design tab, and the transfer Learning block was used for feature generation.

Following this, MobileNetV2 160x160 0.35 was used to train the model, and the accuracy and loss were found to be 86.7% and 0.36, respectively.

Testing and deployment of the model

Testing the generated model was done in Edge Impulse Studio using the Model testing tab, and the model testing accuracy results were found to be 89.47%, respectively.

The model can be deployed using the following command

edge-impulse-linux-runner

The video demonstration is attached here.

The following image shows the experimental setup:

Image 1 shows the experimental setup used for running live classification from the Edge Impulse Studio. As seen in the image Kria development board is connected to a monitor and a webcam. The Kria is also connected to Edge impulse studio, as described earlier. The images are focused in front of the webcam. From the host laptop, live classification is run in Edge Impulse Studio using the Kria KV260. The video demonstration is attached here 1, 2.

The project is made public and can be cloned from Edge Impulse Studio.