AN INTELLIGENT BIRD SCARING SYSTEM - PowerPoint Presentation

AN INTELLIGENT BIRD SCARING SYSTEM FOR CEREAL FARMS WOODS NANCY CHINYERE AHMED MUFUTAU OLAWALE Department of Computer Science University of Ibadan, Nigeria

Introduction Cereal produce like rice, beans are a major source of food for people in most part of the world. Despite the importance of cereal produce like rice in the economy of Nigeria, manual labour still predominates the bird control approach . According to research, it was noted that around 75% of the total farm produce could be destroyed by birds while half of the production costs went into bird scaring. The rise in artificial intelligence and particularly machine learning over the last few years put us at a great feat to solving problems like this in an automated and effective way.

System Design The system is a mobile, portable and self-contained device built on top of Raspberry PI. Several resources such as OpenCV’s Deep Neural Network (DNN) library, TensorFlow object detection API and other third party libraries were utilised

System Design The system essentially consist of a Raspberry PI, a PI camera, a portable speaker, and a bird recognition system built with TensorFlow deployed to the Raspberry PI. The system will read live video stream of a cereal farm via a camera attached to the Raspberry PI. The system passes the video stream read to the raspberry PI.Upon recognition of the bird by the system, the system will hoot random sounds to scare the birds via an attached portable speaker.

Object Detection Model The TensorFlow model zoo is Google’s collection of pre-trained object detection models that have various levels of processing speed and accuracy. The Raspberry Pi has a weak processor and limited RAM (1GB) and for this reason, a model that takes less processing power was needed.The SSD (Single Shot MultiBox Detector) model works fast and requires less computational power as compared to the other model. Though the accuracy is less compared to other models like Faster RCNN Inception.The model has been trained using MSCOCO Dataset which consists of 2.5 million labelled instances in 328, 000 images, containing 90 object types such as “person”, “cat” or “dog”

Object Detection Model The OpenCV DNN module runs much faster than other libraries (SATYA, 2018), and conveniently, it only needs OpenCV in the environment (on the Raspberry Pi). Because of this, the OpenCV DNN module was used to run the SSD model on the Raspberry PI.

Bird Distress Call Bird Distress are sounds that uses bird’s instincts to scare them away from large open spaces . When birds hear these sounds, their natural instinct is to flee the area.For this work, bird sound deterrents as audio files were extracted from a video of bird distress call obtained from Absolute bird control products website (Absolute bird control, 2018).

The Workflow Read Image Frame Convert to blob Feed blob to model Is bird detected ? Output distress call Start No Yes

Proposed final Design

Evaluation Metrics Frame Rate: It is the frequency at which consecutive image called frames are processed by the model and it is calculated by: FPS = No of frames processed/Total time taken Reliability of Detection: To measure this, Recall metric was be used. Recall describes the percentage of relevant objects that are detected with the detector and it is calculated with: Recall = TP/(TP+FN)Reliability of Sound Generation: How reliable the model generate sound on detection of birds.Inference Time: This is the amount of time it takes for the model to detect an object (bird) in a pre-processed frame and it is calculated by: IF = Total time taken/Total no of image frames

Evaluation Data Several video files were obtained from the Internet to test the intelligent bird scarer 50 image frames were extracted from the evaluation data to estimate reliability of detection, inference time and reliability of sound generation. A sample video (about 1 minute) was extracted from the dataset and was used to determine the Frame Rate.

DISCUSSION OF RESULTS

Performance of the Model Frame Rate Performance No of frames Processed Total time taken(seconds) Frame Per second (FPS) 2400 3065.7 0.78 This low performance can be attributed to the low processing power of the Raspberry PI and its limited RAM size (1GB).

Performance of the Model Inference Time The SSDlite Mobilenet model used for the detection purpose has its standard speed of detection (inference) at 27 milliseconds (Rathod, 2017). No. of frames Total Inference Time ( ms ) Average Inference Time ( ms ) 50 43869.212 877.384

Performance of the Model Reliability of Detection Reliability of Sound Generation The reliability of bird distress call sound generation entirely depends on this metric as every time the model predicts accurately, the bird distress call sound is also generated and stopped automatically when no bird is detected. Total True Positive Total False Negative Recall 40 24 0.625

Conclusion An SSD – Mobilenet object detection model was deployed and tested for bird recognition on a Raspberry Pi 3.The level of reliability achieved by the system is reasonably fine given that the system is real-time and that the hardware device used has low computational power.The system could provide an alternative approach to bird scaring in farms that could be deployed to minimize loss due to bird infestation.

Further Work T he performance of the developed system could be greatly improved if more processing power is available on the Raspberry PI. Raspberry PI 4 has just been released. It has 4GB RAM and more computational power. If the model is deployed to this, better performance can be achieved.

Thank YOU!

Questions?