<i>Study of "visualization" method by reconstruction of operation from inertial sensor record of handling tools for cooking</i>

<p>[Purpose] We've been working to understand the movements of a kitchen knife by analyzing operation records from an inertial sensor. However, we encountered two types of signals that obstructed our analysis: sudden noise from the knife's movements and bias from the gravitational pull.</p><p>The sudden noise is caused by the blade's abrupt deceleration when it comes into contact with what's being cut or the cutting board. This results in a brief yet significant peak signal, much larger than what we see from ordinary knife motion.</p><p>The second issue, the bias, comes from the rapidly changing direction of the gravity vector during various knife operations.</p><p>[Method]To overcome these challenges, we tried to recreate the knife's motion in a virtual space, eliminating the interfering signals. To do so, we combined parameters from several digital filters—exponential smoothing, Kalman, median, moving average, and Gaussian—using pre-recorded data of consistent knife operations.</p><p>From this adjusted data, we calculated information on the knife's posture and position, which we then applied to a 3D knife model. The output was a glTF file, suitable for 3D animation.</p><p>[Result] Our approach was effective. The Kalman filter removed the sudden noise, eliminating both angular velocity and acceleration. The median filter used data from 200 points (equivalent to one second of footage) to effectively eliminate the gravity-derived bias. Lastly, we used a complementary filter that employed the angular velocity and acceleration data in conjunction with the gravity vector for correction, allowing us to closely approximate the actual operation.</p>