On-device, real time car fuel lid detection  

We developed a CPU-optimized object detection model for on-device detection of fuel lids on cars.

Background
Fuelmatics is a company dedicated to developing a fully automatic vehicle refueling system. In certain parts of the world, people do not refuel their cars due to either cultural, practical, or legislative reasons. Having staff working to refuel cars is expensive and inefficient – and in deserts in 50C heat, it’s outright dangerous. One way of mitigating these risks and solving the issue is by automating refueling with machines. But to do so, the devices need to quickly identify and locate the lid to the tank.

Solution
The rig detects the fueling lid and uses it as input to move the fueling arm. The lower the prediction latency, the better. The fueling rig has to work detached, without access to an internet connection. And the hardware itself should be as energy-efficient and cheap as possible. This makes the use of GPUs impractical. 

We collected Images of fuel doors from various car models. These were annotated, and several different object detection model candidates were evaluated on inference speed and prediction accuracy. The model was integrated into the central control system of the fueling rig, taking the role of the main detection mechanism, firstly to make sure the car is in the right position, and secondly to guide the fueling arm to the fuel door lid in real-time.

Tools/Tech
The models used were trained using the TensorFlow Object Detection API.  In this case, we resorted to using a pre-trained lightweight object detection model that we fine-tuned. The model was CPU-optimized to be run efficiently on GPU as well as on CPU.