Dynamic Time Warping Distance

Functions
arm_status	arm_dtw_distance_f32 (const arm_matrix_instance_f32 *pDistance, const arm_matrix_instance_q7 *pWindow, arm_matrix_instance_f32 *pDTW, float32_t *distance)
	Dynamic Time Warping distance.
arm_status	arm_dtw_init_window_q7 (const arm_dtw_window windowType, const int32_t windowSize, arm_matrix_instance_q7 *pWindow)
	Window for dynamic time warping computation.
void	arm_dtw_path_f32 (const arm_matrix_instance_f32 *pDTW, int16_t *pPath, uint32_t *pathLength)
	Mapping between query and template.

Description

Dynamic Time Warping Distance.

This is not really a distance since triangular inequality is not respected.

The step pattern used is symmetric2. Future versions of this function will provide more customization options.

Function Documentation

arm_dtw_distance_f32()

arm_status arm_dtw_distance_f32	(	const arm_matrix_instance_f32 *	pDistance,
		const arm_matrix_instance_q7 *	pWindow,
		arm_matrix_instance_f32 *	pDTW,
		float32_t *	distance
	)

Dynamic Time Warping distance.

Parameters

[in]	pDistance	Distance matrix (Query rows * Template columns)
[in]	pWindow	Windowing matrix (can be NULL if no windowing used)
[out]	pDTW	Temporary cost buffer (same size)
[out]	distance	Distance

Returns

ARM_MATH_ARGUMENT_ERROR in case no path can be found with window constraint

Windowing matrix

The windowing matrix is used to impose some constraints on the search for a path. The algorithm will run faster (smaller search path) but may not be able to find a solution.

The distance matrix must be initialized only where the windowing matrix is containing 1. Thus, use of a window also decreases the number of distances which must be computed.

arm_dtw_init_window_q7()

arm_status arm_dtw_init_window_q7	(	const arm_dtw_window	windowType,
		const int32_t	windowSize,
		arm_matrix_instance_q7 *	pWindow
	)

Window for dynamic time warping computation.

Parameters

[in]	windowType	Type of window
[in]	windowSize	Window size
[in,out]	pWindow	Window

Returns

Error if window type not recognized

Windowing matrix

The window matrix will contain 1 for the position which are accepted and 0 for the positions which are rejected.

The input matrix must already contain a buffer and the number of rows (query length) and columns (template length) must be initialized. The function will fill the matrix with 0 and 1.

arm_dtw_path_f32()

void arm_dtw_path_f32	(	const arm_matrix_instance_f32 *	pDTW,
		int16_t *	pPath,
		uint32_t *	pathLength
	)

Mapping between query and template.

Parameters

[in]	pDTW	Cost matrix (Query rows * Template columns)
[out]	pPath	Warping path in cost matrix 2*(nb rows + nb columns)
[out]	pathLength	Length of path in number of points

Warping path

The warping path has length which is at most 2*(query length + template length) in float. 2 because it is a list of coordinates : (query index, template index) coordinate.

The buffer pPath must be big enough to contain the warping path.

pathLength is the number of points in the returned path. The resturned path may be smaller than query + template.