Cleaning Up Our Act with AI

The world is facing a significant increase in waste generation, presenting us with a pressing need to improve our waste management practices. The surge in waste production is driven by a number of factors including population growth, urbanization, industrialization, and consumerism. As a result, landfills are overflowing, oceans are choked with plastic debris, and ecosystems are being compromised. The environmental, social, and economic impacts of this waste crisis are profound and far-reaching.

While recycling has long been touted as a solution to mitigate the environmental impact of waste, it is far from perfect. One major issue is the manual sorting that is required, which is labor-intensive and relies on subjective assessments. Manual sorting also frequently leads to issues with contamination. Contamination occurs when non-recyclable items are mixed with recyclables, rendering the entire batch unsuitable for recycling. For instance, plastic bags thrown into recycling bins can jam recycling machines, necessitating their removal before processing. Consequently, a significant portion of recyclable materials are presently diverted to landfills due to contamination issues.

Moreover, the sheer volume of waste being produced overwhelms manual sorting capabilities. With workers unable to efficiently sift through all waste streams and remove contaminants, automated detection solutions are urgently needed to enhance recycling processes. Such a solution has just been proposed by a team at NVIDIA. They have developed a prototype system that is capable of analyzing imagery from the hopper of a waste collection truck to detect a particularly troublesome and common source of contamination — plastic bags.

This was made possible by leveraging an NVIDIA Jetson TX2 edge computing device. Jetsons are equipped with NVIDIA GPUs and are engineered to operate in low power consumption modes that make them suitable for running advanced machine learning algorithms in mobile applications. A camera, of the sort that is frequently deployed on waste collection trucks for manual inspections, was connected to the Jetson to allow for automated image analysis.

To pick regions of contamination out of an image, an object detection algorithm was needed. In this case, the team chose to use a YOLOv4 model, which they retrained on a custom waste contamination dataset. This dataset, called the Remondis Contamination Dataset, was produced in collaboration with a partner in the waste management field, and contains images from the hopper of a collection truck, with annotation specifying the location of any plastic bags that are present.

On the Jetson TX2, the object detection algorithm ran at a blazing 24.8 frames per second, which should be faster than necessary for this application. With a mean average precision of 63 percent, the proof of concept device showed its utility, and after retraining on additional images collected in the field, that precision increased by an additional 10 percent. With a bit more work, the predictions should be good enough for a production system.

Looking ahead, further development will be needed to make this contamination detector more useful. While it can detect plastic bags with adequate levels of accuracy, there are other types of contaminants that will also need to be detected. Moreover, once detected, some action will need to be taken to fully address the problem. Once these enhancements have been made, the team suggests that additional cameras could be installed on the trucks to solve other problems, like by detecting potholes or roadside trash, for example.

Further details are available in the team's blog post about the project for those that want to dig deeper or build their own edge object detection solution.