The foundation of the
system is its visual model, which Cheng trained on a custom mosquito dataset. To do that, he relied on a DSLR camera with a high-magnification zoom lens, capturing detailed images of mosquitoes for training data. That same camera setup serves as both the training tool and the main sensor for detecting mosquitoes.
Cheng said the process of collecting this data left him with "countless mosquito bites all over my body," a reminder that even highly technical projects sometimes require getting uncomfortably up close and personal with the real world. The images he gathered were annotated and used to train a deep learning model to identify mosquitoes in flight.
Cheng noted that the process "really put my graphics card through its paces" as the system processed the data to improve detection accuracy. By the end, he said the model's performance was "quite good," suggesting it could reliably pick out mosquitoes from background noise.
With detection in place, Cheng turned to the response system. He integrated a laser that was calibrated, in his words, to "instantly turn mosquitoes into roasted ones." The laser is mounted on a high-precision industrial rotary stage, enabling it to move quickly and accurately while tracking targets identified by the vision system.
The result is a closed-loop setup: the camera identifies a mosquito, the model confirms it, and the hardware adjusts in real time to aim and fire. Unlike traditional bug zappers, which rely on passive attraction, this system actively tracks and engages individual insects.
