Let's Play Waste Sorting!

The idea of waste sorting is awesome. The problem is that people don't do it.

This project aims to educate and encourage people to sort their waste better while helping preserve nature, lowering their garbage collection bill, and making service workers' jobs easier.

This is a general overview of the project, which is still in development. There are many challenges involved with detecting sorting trash, which are still being worked on. We are also experiencing very poor communication from our local waste management companies, which makes gathering of information more difficult and time consuming. This is probably partly due to the COVID-19 situation as many people are working from home and don't have the time or resources to aid us.

I wanted to share the idea and my envisioning of this system's implementation. And just maybe, someone will find this an interesting read.

There are Pay-As-You-Throw (PAYT) systems implemented all over the world

PAYT systems have been used in some way, shape or form for over 100 years. They are present in USA and Canada, a lot of European countries and some Asian countries. Different places have different implementations of the PAYT system, rewarding schemes and different enforcing policies, but generally, the impact on recycling rates has been evaluated as positive overall.

In most cases, the system works either by weighing the amount of garbage in a bin or by using standard-sized bags and charging a per-bag fee upon collection. This is a proven system that works, but has room for improvement, especially by adding the ability to empower waste management facilities to raise awareness about recycling within their area of influence.

This is different

The proposed system uses a camera mounted on a garbage collection truck to identify incorrectly sorted waste in containers for recyclables. By tagging garbage bins with QR codes, identification is also carried out by the camera. The system can be used in conjunction with current systems relying on weight measurements, to produce more in-depth data on the state of waste sorting in a specific geographical area. This data can then be used by waste management companies to dynamically adjust the collection bill and get insight into specific incorrectly sorted items, enabling them to raise awareness among communities that they service.

What about privacy?

One of the biggest challenges the project could face is the issue of privacy, namely, people not wanting to have pictures of their trash taken and stored on some company's server. The system tackles this issue by using on-board image processing and only transmitting the analyzed data to the server, and deleting the image immediately afterwards. This ensures that the image never leaves the physical ecosystem mounted on the garbage collection truck.

When users get feedback on the mobile app, the comparison data is completely anonymized. There is no single user data given to any other user. The user's data is shown to them in comparison to different wider geographical areas, like their district, city, state, country or even globally, instead.

GPS coordinates of detected bins are also processed on board which can help detect placement errors, thefts or 'bin switches', which could fraudulently manipulate the system. The waste management company gets notified that something is wrong at the given location and a person can be sent to go check it out.

While deploying the system, the opt-in model should be used. This generates a user base of early adopters who are more tolerant to hiccups in the system and are more willing to give feedback about them to the deployer. If sorting waste correctly, they should see a decrease in their waste collection bill, and adoption would rise by word of mouth.

The key thing when deploying such a network is transparency. Users should be unambiguously told how the system works, how their data is processed, what amount of that data actually gets stored and processed on the server, and what data is displayed to them and to other users. The contents of a garbage bin can be a delicate topic for some people, and this is the only way to gain people's trust and actually scale beyond enthusiast users.

How does it work?

The system is comprised of three main components:

• the on-board unit,
• the server,
• the mobile app.

The on-board unit is run by a Raspberry Pi in the cabin of the truck, which receives images from a Wi-Fi-enabled security camera, mounted on the back of the truck. When a garbage bin with a QR code on top is put on the truck's arms, it gets identified by the camera and the GPS coordinates get collected from a GPS unit, connected to the Raspberry Pi. When the bin is emptied in the preliminary stage of the truck's collection and compression mechanism, another picture is taken, which captures the contents of the bin. The image is then transmitted via Wi-Fi to the Raspberry Pi which uses an algorithm in conjunction with the Coral AI accelerator to assess whether there are any incorrectly sorted items in there. The gathered data is then transmitted to the server using an LTE CAT-M module connected to the Raspberry Pi.

Theserver is responsible for storing the received data in a database and then, based on it:

• determine the cost of the collection service for each instance of data,
• doing statistical analysis on the data for the user and differently sized geographical areas around them,
• determining trends for the user and the aforementioned geographical areas,
• comparing those to other users and geographical areas within the country of deployment, and
• pushing relevant data to the user's mobile app.

The mobile app's main goal is to give users feedback on how their sorting is improving or worsening and encouraging them to do better. It has a couple of features:

• a general overview: the user's sorting trends, comparison to others,
• statistical analysis of the user's sorting over time,
• a display of bill amount over time and notifications when improving or worsening significantly with recommendations on how to do better,
• the ability to pay unpaid bills within the app or with an external payment service,
• a section with tips on what items to put in which recycling bin.

It should mainly be focused on making the user feel good about their recycling work with competing against others with the possibility of having a financial gain because of it. In the future, we are thinking of adding certain gamification aspects to the app, to make it even more interactive and appealing to users and to encourage them to try to recycle better.

The elephant in the room

Where's the detection code?

Until the date of posting, we have not yet developed an algorithm, able to successfully determine the 'quality of recycling'. We haven't really decided what the determining factor would be. We cannot search images for aluminum cans, since we cannot tell what was in the cans, but we have to teach our algorithm to detect labels and shapes in order to determine whether they contained juice or motor oil. And even if we could, we still aren't sure this is the best way to do it. We currently have very little information from our local waste management companies on what kind of waste is most often wrongly sorted and what their suggestions are.

Another problem we thought of is what do we do if the garbage is stored in a bag? Do we ignore it? Do we enforce out users to empty their bags into the container? Currently, we have decided to only work on the 'recyclables' bin, which takes most of food packaging, plastic bags, plastic bottles and similar things, but should not contain motor oil, hair spray, different types of paint, etc. We picked this type of bin because there is usually the least amount of bagged items in it and have found out that about 30% of them are incorrectly sorted.

There are certain other problems, which mostly originate from lack of information, which we are hoping to acquire in the future.

So... What has been done so far?

While most of our time was spent on research, we also started working on some technical aspects of the project. We created an identification script which uses OpenCV to detect QR codes, printed onto waste bins. It also saves the image to the disk, if this is the first occurrence of a certain QR code. It currently uses a USB webcam as a proof-of-concept, but we can change it later to use an IP camera stream instead.

We also created a prototype Android app which enables viewing stats and checking finances or paying bills. The design is still left default and is subject to some change in the future. It is a purely functional display prototype of the app. We are hoping that it will look something like this in the first iteration:

In conclusion

We think that adopting this kind of system will help everyone involved in the waste management process. Users would see financial and social encouragement and make a game out of it, waste management facilities would have to do less post-processing on their intakes which would lead to lesser cost and workload of sorting machines and personnel, and nature would benefit through raised awareness.

In the future, we also need a person with experience in machine learning, more specifically object detection and classification algorithms, who could give us more insight into how to tackle the problem altogether and help us solve it.