Technology

Inspiration

After the beach cleanup activates, we found that picking up microplastics by hand is ineffective and tedious, therefore we thought of an idea to carry out the procedures with a robotic car.

Aim

We aim to develop a prototype of an automated microplastic collecting car which collects sand, and filters out the microplastics all by its own for microplastics collection and educational purposes.

Basic Car Components

The backbone of the car is mainly built with beams of a mBot Ultimate 2.0 kit. The main board of the car is MegaPi and consists of 4 DC Encoder Motor controlled by 4 MegaPi Encoder/DC Driver V1 separately: 2 for controlling the track wheels, 1 for the rotation of the specialized circular cylinder and 1 for the rotation of the sieve component. A Gyro sensor is also attached to the cylinder for monitoring its attitude.

Design

The NGO, The Big Microplastic Survey, suggests that we take sand samples of 50 cm³, 5 cm (width) x 5 cm (width) x 2 cm (depth). Our design must meet the requirements of collecting sand volume and extract microplastics within them. Thus, the circular cylinder of the car is designed in a combined shape of circle and ellipse and features a shovel-like opening. The cylinder is installed close to the horizon with the protruding part designed specifically to be around 9.31 cm² with the height of 5.37cm which is calculated to be 50 cm³ in total to meet our requirements.

The robotic car conducts a repeating mechanical process of collecting microplastic every 2 meters along the coastline of the beach. (Distance varies with the data wanted) Then the collected microplastic will be stored in a sealed box and brought back to lab for analysis and data collection.

Mechanism

The circular cylinder in the car is programmed to rotate 360˚ to effectively scoop up the sand, and water can be subsequently pumped in. When the cylinder vibrates where the motor rotates back and forth for, the specialized physical patterns inside the cylinder facilitates the blending procedure of the sand and water. This creates an environment for the microplastics to float on the water surface as a result of their density difference. Following up, the car delays 30 seconds for the microplastic to fully emerge and the sieve component rotates 180˚ to appear in front of the cylinder. The cylinder then tilts at a 135-degree angle, facilitating the pouring of the water's surface layer, along with microplastics, into the sieve for collection. After that, the sieve component rotates 180˚ back while the cylinder rotates 45˚ and vibrates to discharge the sand. Then the cylinder rotates 180˚ to ensure a continuous loop for the entire process.

Further advancements/Phases

We have divided the development of our microplastic collecting car into three phases.
Phase I
The first phase is functioning the basic mechanism mentioned above to collect microplastics.
Phase II
The second phase is the addition of a GPS function to the machine, which means it can collect microplastics fully automated as it knows its position. It will then follow the coastline, where most microplastics are located to collect most samples.
Phase III
The third phase is about analysing the data. We planned to introduce an object detection AI using Mask R-CNN to mask and count the microplastics under microscopic images. With slight modifications, the data we obtain can be used to build a 3D model of the microplastic distribution on the beach, which could be used for further investigations and researches.

Powerpoint

---

As mentioned above, we planned to introduce an object detection AI to mask and count the microplastics under microscopic images.

The detection algorithm we’re using is Mask-RCNN[1], which is an object detection algorithm developed by Facebook. The algorithm extends the concept of Faster-RCNN, which only creates bounding boxes around detected objects. We decided to use Detectron2[2] as a framework to create a model.

As for our first trial, we have obtained our microplastic images from Unit For Data Science[3]. Then we utilized a Computer Vision Annotation Tool(CVAT)[4] to annotate the microplastic images in 3 classes(microplastic bead, microplastic fiber and microplastic fragment) and generate our training dataset in a COCO format.

For our preliminary test, our training and validation dataset contains 159 images in total, and are split in an 8:2 ratio, where there are 127 and 32 images for training and validation. We trained the models in Google Colaboratory with T4 GPU, for 1000 iterations.
We trained the model in different baselines from the Detectron Model Zoo[5], which includes R101-FPN, R50-FPN.

R_50_FPN bbox AP

R_50_FPN segm AP

R_101_FPN bbox AP

R_101_FPN segm AP

By comparing the Average Precisions(AP) on the results of R_50_FPN and R_101_FPN, we discovered that R_101_RPN have an overall better Average Precision than R_50_FPN(+2.10, +1.54)

R_50_FPN

R_101_FPN

From the training curves shown above, the loss of R_50_FPN and R_101_FPN both reached stability in around 600 iterations

Image results

R_50_FPN

R_101_FPN

From the mapping results above, we can see R_101_FPN has obtained an overall better performance than R_50_FPN. However, there are situations where the AI cannot differentiate between microplastic beads and fiber, and sometimes misrecognize other substances as microplastics. This is because our small scaled dataset is entirely not enough to achieve maximum performance.

As a preliminary test, we noticed R_101_FPN has an overall better performance in mapping microplastic which we will further investigate on this baseline for our future AI model. Furthermore, we may need to create a larger microplastic dataset and apply data augmentation to the dataset for a better performance.

References:

[1]https://arxiv.org/abs/1703.06870

[2]https://ai.meta.com/tools/detectron2

[3]https://github.com/UnitForDataScience/Microplastic_Image_Analysis

[4]https://www.cvat.ai/

[5]https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md