About the Client
The client is an Israeli startup that develops computer-vision-based software for precision livestock monitoring and growth optimization.
Agriculture in Israel is a highly developed industry. Because the geography of the country is not naturally suited for it, Israel was forced to get creative around the efficiency of natural resources. Today, the country is a leader in agricultural technologies and livestock. Being a part of the agriculture sector, livestock is of major economic and social importance.
The client wanted to build a solution that would give farmers advice on how to achieve the optimal livestock population and automatically administer the stock in real-time. The solution in mind would also have to provide automatic monitoring, tracking, and weighing the of the stock, thus speeding up the measurement process, facilitating human labor, and decreasing farm costs.
Quantum built a real-time, cloud-based pig farm monitoring platform that detected, classified, and quantified livestock.
- algorithms for data preparation, munging, and image processing
- optimization algorithms for parameter estimation
- algorithms for results prediction
The solution was a web-based service that connects to farm servers, runs the algorithm to classify animals in real-time and receives a live stream from the cameras on the farm.
Data is wirelessly sent to the cloud, where image processing starts: detecting, identifying and weighing pigs. All results are saved to the database. This approach is perfect for farmers who no longer have to wait for hours or an entire day to measure and compare data about their livestock.
The solution can also monitor deviations in animal populations and help manage the growth process more accurately, saving money and reducing farm losses.
The Quantum team used machine learning to predict, categorize, classify and measure animals from the given datasets of the live-stream from cameras on farms.
The solution pipeline was the following:
- detection and identification of ID tags
- detection of pigs using instance segmentation NN
- binding detected tags and pigs; results filtering and choosing the best one
- identification of pig body parts (body, head, legs) using semantic segmentation NN
- prediction of pigs’ weight using nonlinear regression with an output of the previous step as the model input
Let's discuss your idea!
Quantum had to create a fast and scalable web service with exceptional scalability. The team used Python as it allows developing asynchronous APIs easily and has a full spectrum of ML, DL, and CV libraries.
The whole pipeline allows us to process 1 frame per second, a prototype was developed using FullHD cameras with a wide view angle. Processing was executed at the cloud using AWS SageMaker and Tensorflow serving. The GPU used for processing is NVidia M60. While developing a solution state-of-the-art model was used for each task (Mask-R CNN and customized UNet). Segmenting separately body, head and tail decreased error of weighing by 5%.