Counting Broiler from Poultry House Video Using Faster Region-based Convolutional Neural Networks

In a poultry house, livestock object detection is essential for farmers to prepare the amount of food. Counting broiler can be done through video monitoring placed at the top. If video processing is done automatically, it will collide with memory, especially with many frames. Image approach can be taken considering the broiler is in a poultry house or closed room, so that one frame per second processing is quite representative. In addition, the observed broilers are considered to have a similar shape, and their identity is ignored. Therefore, this study proposed using the Faster Region-based Convolutional Neural Network (Faster R-CNN) method for broiler livestock. Faster R-CNN is one of the models of object detection that can be used to count the number of objects. In Faster RCNN, choosing the proper architecture and configuration must be considered to find good performance. The experiment showed that using ResNet-101, Faster R-CNN achieves the best accuracy than other famous CNN architecture, such as ResNet-50 and ResNeXt-101, with an average precision of 50% achieving 92.8% and average precision of 75%, achieving 78.1%. The ResNet-101 architecture has more convolution depth. The depth causes ResNet-101 to do better feature extraction even though ResNeXt-101 has more depth convolution, but it made ResNeXt-101 too heavy to do feature extraction, making ResNeXt-101 have the largest size model. However, the Faster R-CNN model using ResNet-101 is better implemented for real-time detection.