Waste Management is the leading provider of comprehensive waste management services in North America and saw the need for a detection system that would protect on-vehicle riders. Rear loader residential collection vehicles require two passengers to ride on the back of the vehicle as waste is collected. Those riders risk severe injury, even death, from the occasional rear-end collisions with other passenger cars. Currently, there is no detection system to alert the riders to potential oncoming accidents, besides the riders and driver being alert and aware. Creating an automated detection system for the oncoming vehicles that pose risk for the riders could reduce injuries and save lives.
To provide a solution that would be easy to implement and keep riders safe, Waste Management conceived of and envisioned an edge device and hardware system that could be embedded onto trucks to alert riders when oncoming traffic poses a risk, and engaged KUNGFU.AI to help develop a platform for a computer vision model that would identify oncoming risks and trigger an alert. The device is powered by a USB battery and features Edge TPUs that allow for high-performance machine learning computing on low-power devices. We custom-built an ensemble of computer vision models, running on edge devices and deployed on Google Cloud Platform. We leveraged state-of-the-art detection and segmentation frameworks to bifurcate trailing vehicles from other non-threatening objects. In addition, we built a threat prediction model that calculates speed, velocity, and distance to make a threat prediction for identified oncoming vehicles. If the vehicle was deemed a potential threat, an embedded sensor would alert riders to brace for impact.
The Waste Management Early Alert Rear Safety Device has been tested in the field and successfully alerted on-vehicle riders to potential oncoming danger. Before deploying on vehicles, parking lot testing revealed that detection took 0.05 second per frame, with tracking and scoring taking 0.05 second per frame. The performed testing was highly accurate with no false negatives reported.
Segmentation model that can understand the difference between people, objects, and moving vehicles.
Demonstration of a model calculating the threat of an oncoming vehicle.
Waste Management is the leading provider of comprehensive waste management services in North America and saw the need for a detection system that would protect on-vehicle riders. Rear loader residential collection vehicles require two passengers to ride on the back of the vehicle as waste is collected. Those riders risk severe injury, even death, from the occasional rear-end collisions with other passenger cars. Currently, there is no detection system to alert the riders to potential oncoming accidents, besides the riders and driver being alert and aware. Creating an automated detection system for the oncoming vehicles that pose risk for the riders could reduce injuries and save lives.
To provide a solution that would be easy to implement and keep riders safe, Waste Management conceived of and envisioned an edge device and hardware system that could be embedded onto trucks to alert riders when oncoming traffic poses a risk, and engaged KUNGFU.AI to help develop a platform for a computer vision model that would identify oncoming risks and trigger an alert. The device is powered by a USB battery and features Edge TPUs that allow for high-performance machine learning computing on low-power devices. We custom-built an ensemble of computer vision models, running on edge devices and deployed on Google Cloud Platform. We leveraged state-of-the-art detection and segmentation frameworks to bifurcate trailing vehicles from other non-threatening objects. In addition, we built a threat prediction model that calculates speed, velocity, and distance to make a threat prediction for identified oncoming vehicles. If the vehicle was deemed a potential threat, an embedded sensor would alert riders to brace for impact.
The Waste Management Early Alert Rear Safety Device has been tested in the field and successfully alerted on-vehicle riders to potential oncoming danger. Before deploying on vehicles, parking lot testing revealed that detection took 0.05 second per frame, with tracking and scoring taking 0.05 second per frame. The performed testing was highly accurate with no false negatives reported.
Segmentation model that can understand the difference between people, objects, and moving vehicles.
Demonstration of a model calculating the threat of an oncoming vehicle.
