How is the kalman filter initialized?

[prass-anyvision]

What are these factors in :-

        std = [
            2 * self._std_weight_position * measurement[3],
            2 * self._std_weight_position * measurement[3],
            1e-2,
            2 * self._std_weight_position * measurement[3],
            10 * self._std_weight_velocity * measurement[3],
            10 * self._std_weight_velocity * measurement[3],
            1e-5,
            10 * self._std_weight_velocity * measurement[3]]

https://github.com/nwojke/deep_sort/blob/280b8bdb255f223813ff4a8679f3e1321b08cdfc/deep_sort/kalman_filter.py#L76

[studentbrad]

Those factors initialize the coveriance matrix using scalar multiples of the bounding box height. A solid way of initializing the covariance matrix given that no depth information is included in an image. Because we don't expect the aspect ratio to change much we use small numbers like 1e-2 and 1e-5 for the aspect ratio velocity. You can read about this assumption in the publication. https://arxiv.org/abs/1703.07402