Kalman filterI've done this for five yearsAutopilotIt's only in the state estimates that we really get through. I've seen a lot of people go into that array equation, but it's not like that. It is essentially an updated set of "half letters, half doubt" rules: sensor measurements are unreliable and model predictions are biased, so that they can be combined with the best weights and can be estimated in a state more than any single source. Ten minutes to finish, three years to step on the project pit, and we'll talk about it in conjunction with my actual project。
How does kalman filter understand
The core is two steps away. In the prediction phase, you take one step with a state-of-the-art and control input, and you build up the matrix -- "i don't see it, so the uncertainty rises." update, new measurements. Count oneKalman gainsIt's essentially a self-adaptation weight: measuring noise is big, it's small, you believe in models, it's big, it's good, you do. It's just a scaling logic that doesn't need a high instinct to hold it。
What really makes a big head is how the process noise queen works. My early drone program filled with q's full head, and the filters were either locked or convulsedProportional relationsIt's life. It's not that important。
Where's the kalman filter
The classic scenario is gps and imu integration. The gps jumps in one position for a few seconds, and the imu sews the imu with 100 hz accelerator and angular velocity, with a timescale of several dozen times, and the kalman filter sews the successive imu predictions and gps thinness measurements. I've done a project where imu points eight meters in 10 minutes, and it's filtered down to a dozen centimetres, and it's a complete two things。
Don't just stareNavigationI don't know。RobotSlam position estimates, industrial power retroactivity observations, and even quantitative transactions with ekf estimates (e. G. Market volatility), i have a friend in finance who says it's better than a straight historical average, which is the use of tools outside the border。
Why can't kalman filter
The linear assumption is the original offence. Real systems are mostly non-linear, and hard-standard kalman gains are crooked. The project is generally based on ekf or ukf, but even using ukf to measure a jump in or drop a package, the bad points can be delayed for several seconds. I added it laterMa's range, the point of the super-threshold drops directly and the output stabilizes more than one scale。
There's the beginning. The p matrix is too big, the first dozens of steps are "learning" and it's a piece; it's too small for filters to believe the model is being updated and slow. Without silver bullets, a reasonable starting point for offline data is better than anything。
The tools are smooth, the bottom noise model doesn't match, and it's garbage coming in and out. Next time someone asks kalman about the filter, i usually say, ten minutes, three years。
