Development

Our Progress

2022

June 15

First Presentation

It was interesting!

June 14

To move or not to move

Our team faced a dilemma on whether we should allow the camera mounted on our turret to move along with the laser whenever it’s detecting and firing the laser at the birds. After consulting Ms Sarah, we decided to let the camera move along with the laser as doing otherwise would drastically increase the complexity of aiming the laser at our subject.

June 1

Second Meeting with Dr Ho (another interesting idea!)

This time, we thought of using lasers to scare the birds away. Lasers are much more accurate than salt pellets, whose flight trajectories can easily be altered by any wind. Besides, low-powered lasers do not harm birds and are proven effective. Thankfully, our idea was approved this time!

May 25

First Meeting with Dr Ho

We shared our brilliant idea of scaring birds from food centers by firing salt pellets at them! However, we were advised to find some other means of deterring birds for obvious reasons…

Hurdle 1: Choice of object detection algorithmNow that we had the idea for our device, the next question was where it could be used. In society, there are several places where such a device could be deployed. The most obvious place would be food centers where birds constantly feed on leftover food at tray return counters before the cleaners could clear the used utensils. However, to avoid running the risk of accidentally injuring others with the laser, we decided that the device should be used in low-traffic areas like farmlands where farmers constantly face the problem of pest birds feeding on their harvest.

In order to deter birds in large areas like farmlands, the object detection algorithm used to detect the birds must first be very accurate. At first, we managed to find a convolutional neural network model called YOLOv3 (You Only Look Once, Version 3) that was fairly accurate in detecting small birds in the frame (typical for birds that feed in farmlands since they are far away). However, a trade-off exists between the accuracy of a model and the computational cost incurred when running it. Although the algorithm worked well on our laptops, it performed very poorly on the Jetson Nano we were using as the Jetson Nano has relatively weaker processing capabilities. As such, we were only limited to using lightweight models with mediocre accuracy in detecting birds. This also meant that we had no choice but to restrict our project to mainly deterring birds that are close to the camera, such as in food centers. Eventually, we settled on an object detection called MobileNetV2 that is decently accurate yet lightweight enough to be used in mobile phones with low processing capabilities. Here is a quick video to illustrate the difference in performance between both algorithms on the Jetson Nano

Initially, we planned to use 4 cameras on our device to allow for complete coverage of its surrounding area.

Initially, we planned to use the Arduino. However for the final product, the Jetson Nano was used.

However, it would introduce a lot more complexity to our project. Firstly, we would need to find the optimal way to turn the laser towards the bird (it could either turn clockwise or anti-clockwise), which is further complicated when the device is surrounded by birds. Doing so might also be overly computationally expensive. Hence, for now, we decided to stick with a single camera.
Another dilemma we faced was in deciding whether the camera should move with the laser diode. In our original design, the camera was affixed to the base of the camera and does not rotate with the laser diode. However, doing so would drastically increase the complexity of aiming the laser at the subject. Therefore, we changed the design of our prototype such that the camera now moves with the laser diode.