/* * Copyright (c) 2016-2019 ARM Limited. * * SPDX-License-Identifier: MIT * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to * deal in the Software without restriction, including without limitation the * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in all * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "arm_compute/core/NEON/kernels/NELKTrackerKernel.h" #include "arm_compute/core/AccessWindowStatic.h" #include "arm_compute/core/Coordinates.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" #include #include using namespace arm_compute; /** Constants used for Lucas-Kanade Algorithm */ constexpr int W_BITS = 14; constexpr float D0 = 1 << W_BITS; constexpr float DETERMINANT_THRESHOLD = 1.0e-07f; // Threshold for the determinant. Used for lost tracking criteria constexpr float EIGENVALUE_THRESHOLD = 1.0e-04f; // Thresholds for minimum eigenvalue. Used for lost tracking criteria constexpr float FLT_SCALE = 1.0f / (1 << 20); namespace { enum class BilinearInterpolation { BILINEAR_OLD_NEW, BILINEAR_SCHARR }; template constexpr int INT_ROUND(T x, int n) { return (x + (1 << (n - 1))) >> n; } template inline int get_pixel(const ITensor *tensor, int xi, int yi, int iw00, int iw01, int iw10, int iw11, int scale) { const auto px00 = *reinterpret_cast(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi, yi))); const auto px01 = *reinterpret_cast(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi + 1, yi))); const auto px10 = *reinterpret_cast(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi, yi + 1))); const auto px11 = *reinterpret_cast(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi + 1, yi + 1))); return INT_ROUND(px00 * iw00 + px01 * iw01 + px10 * iw10 + px11 * iw11, scale); } inline int32x4_t compute_bilinear_interpolation(int16x8_t top_row, int16x8_t bottom_row, int16x4_t w00, int16x4_t w01, int16x4_t w10, int16x4_t w11, int32x4_t shift) { // Get the left column of upper row const int16x4_t px00 = vget_low_s16(top_row); // Get the right column of upper row const int16x4_t px01 = vext_s16(px00, vget_high_s16(top_row), 1); // Get the left column of lower row const int16x4_t px10 = vget_low_s16(bottom_row); // Get the right column of right row const int16x4_t px11 = vext_s16(px10, vget_high_s16(bottom_row), 1); // Apply the bilinear filter return vqrshlq_s32(vmull_s16(px00, w00) + vmull_s16(px01, w01) + vmull_s16(px10, w10) + vmull_s16(px11, w11), shift); } } // namespace void NELKTrackerKernel::init_keypoints(int start, int end) { if(_level == _num_levels - 1) { const float level_scale = pow(_pyramid_scale, _level); for(int i = start; i < end; ++i) { _old_points_internal->at(i).x = _old_points->at(i).x * level_scale; _old_points_internal->at(i).y = _old_points->at(i).y * level_scale; _old_points_internal->at(i).tracking_status = true; NELKInternalKeypoint keypoint_to_track; if(_use_initial_estimate) { keypoint_to_track.x = _new_points_estimates->at(i).x * level_scale; keypoint_to_track.y = _new_points_estimates->at(i).y * level_scale; keypoint_to_track.tracking_status = (_new_points_estimates->at(i).tracking_status == 1); } else { keypoint_to_track.x = _old_points_internal->at(i).x; keypoint_to_track.y = _old_points_internal->at(i).y; keypoint_to_track.tracking_status = true; } _new_points_internal->at(i) = keypoint_to_track; } } else { for(int i = start; i < end; ++i) { _old_points_internal->at(i).x /= _pyramid_scale; _old_points_internal->at(i).y /= _pyramid_scale; _new_points_internal->at(i).x /= _pyramid_scale; _new_points_internal->at(i).y /= _pyramid_scale; } } } std::tuple NELKTrackerKernel::compute_spatial_gradient_matrix(const NELKInternalKeypoint &keypoint, int32_t *bilinear_ix, int32_t *bilinear_iy) { int iA11 = 0; int iA12 = 0; int iA22 = 0; int32x4_t nA11 = vdupq_n_s32(0); int32x4_t nA12 = vdupq_n_s32(0); int32x4_t nA22 = vdupq_n_s32(0); float keypoint_int_x = 0; float keypoint_int_y = 0; const float wx = std::modf(keypoint.x, &keypoint_int_x); const float wy = std::modf(keypoint.y, &keypoint_int_y); const int iw00 = roundf((1.0f - wx) * (1.0f - wy) * D0); const int iw01 = roundf(wx * (1.0f - wy) * D0); const int iw10 = roundf((1.0f - wx) * wy * D0); const int iw11 = D0 - iw00 - iw01 - iw10; const int16x4_t nw00 = vdup_n_s16(iw00); const int16x4_t nw01 = vdup_n_s16(iw01); const int16x4_t nw10 = vdup_n_s16(iw10); const int16x4_t nw11 = vdup_n_s16(iw11); // Convert stride from uint_t* to int16_t* const size_t row_stride = _old_scharr_gx->info()->strides_in_bytes()[1] / 2; const Coordinates top_left_window_corner(static_cast(keypoint_int_x) - _window_dimension / 2, static_cast(keypoint_int_y) - _window_dimension / 2); auto idx = reinterpret_cast(_old_scharr_gx->buffer() + _old_scharr_gx->info()->offset_element_in_bytes(top_left_window_corner)); auto idy = reinterpret_cast(_old_scharr_gy->buffer() + _old_scharr_gy->info()->offset_element_in_bytes(top_left_window_corner)); static const int32x4_t nshifter_scharr = vdupq_n_s32(-W_BITS); for(int ky = 0; ky < _window_dimension; ++ky, idx += row_stride, idy += row_stride) { int kx = 0; // Calculate elements in blocks of four as long as possible for(; kx <= _window_dimension - 4; kx += 4) { // Interpolation X const int16x8_t ndx_row1 = vld1q_s16(idx + kx); const int16x8_t ndx_row2 = vld1q_s16(idx + kx + row_stride); const int32x4_t nxval = compute_bilinear_interpolation(ndx_row1, ndx_row2, nw00, nw01, nw10, nw11, nshifter_scharr); // Interpolation Y const int16x8_t ndy_row1 = vld1q_s16(idy + kx); const int16x8_t ndy_row2 = vld1q_s16(idy + kx + row_stride); const int32x4_t nyval = compute_bilinear_interpolation(ndy_row1, ndy_row2, nw00, nw01, nw10, nw11, nshifter_scharr); // Store the intermediate data so that we don't need to recalculate them in later stage vst1q_s32(bilinear_ix + kx + ky * _window_dimension, nxval); vst1q_s32(bilinear_iy + kx + ky * _window_dimension, nyval); // Accumulate Ix^2 nA11 = vmlaq_s32(nA11, nxval, nxval); // Accumulate Ix * Iy nA12 = vmlaq_s32(nA12, nxval, nyval); // Accumulate Iy^2 nA22 = vmlaq_s32(nA22, nyval, nyval); } // Calculate the leftover elements for(; kx < _window_dimension; ++kx) { const int32_t ixval = get_pixel(_old_scharr_gx, top_left_window_corner.x() + kx, top_left_window_corner.y() + ky, iw00, iw01, iw10, iw11, W_BITS); const int32_t iyval = get_pixel(_old_scharr_gy, top_left_window_corner.x() + kx, top_left_window_corner.y() + ky, iw00, iw01, iw10, iw11, W_BITS); iA11 += ixval * ixval; iA12 += ixval * iyval; iA22 += iyval * iyval; bilinear_ix[kx + ky * _window_dimension] = ixval; bilinear_iy[kx + ky * _window_dimension] = iyval; } } iA11 += vgetq_lane_s32(nA11, 0) + vgetq_lane_s32(nA11, 1) + vgetq_lane_s32(nA11, 2) + vgetq_lane_s32(nA11, 3); iA12 += vgetq_lane_s32(nA12, 0) + vgetq_lane_s32(nA12, 1) + vgetq_lane_s32(nA12, 2) + vgetq_lane_s32(nA12, 3); iA22 += vgetq_lane_s32(nA22, 0) + vgetq_lane_s32(nA22, 1) + vgetq_lane_s32(nA22, 2) + vgetq_lane_s32(nA22, 3); return std::make_tuple(iA11, iA12, iA22); } std::pair NELKTrackerKernel::compute_image_mismatch_vector(const NELKInternalKeypoint &old_keypoint, const NELKInternalKeypoint &new_keypoint, const int32_t *bilinear_ix, const int32_t *bilinear_iy) { int ib1 = 0; int ib2 = 0; int32x4_t nb1 = vdupq_n_s32(0); int32x4_t nb2 = vdupq_n_s32(0); // Compute weights for the old keypoint float old_keypoint_int_x = 0; float old_keypoint_int_y = 0; const float old_wx = std::modf(old_keypoint.x, &old_keypoint_int_x); const float old_wy = std::modf(old_keypoint.y, &old_keypoint_int_y); const int iw00_old = roundf((1.0f - old_wx) * (1.0f - old_wy) * D0); const int iw01_old = roundf(old_wx * (1.0f - old_wy) * D0); const int iw10_old = roundf((1.0f - old_wx) * old_wy * D0); const int iw11_old = D0 - iw00_old - iw01_old - iw10_old; const int16x4_t nw00_old = vdup_n_s16(iw00_old); const int16x4_t nw01_old = vdup_n_s16(iw01_old); const int16x4_t nw10_old = vdup_n_s16(iw10_old); const int16x4_t nw11_old = vdup_n_s16(iw11_old); // Compute weights for the new keypoint float new_keypoint_int_x = 0; float new_keypoint_int_y = 0; const float new_wx = std::modf(new_keypoint.x, &new_keypoint_int_x); const float new_wy = std::modf(new_keypoint.y, &new_keypoint_int_y); const int iw00_new = roundf((1.0f - new_wx) * (1.0f - new_wy) * D0); const int iw01_new = roundf(new_wx * (1.0f - new_wy) * D0); const int iw10_new = roundf((1.0f - new_wx) * new_wy * D0); const int iw11_new = D0 - iw00_new - iw01_new - iw10_new; const int16x4_t nw00_new = vdup_n_s16(iw00_new); const int16x4_t nw01_new = vdup_n_s16(iw01_new); const int16x4_t nw10_new = vdup_n_s16(iw10_new); const int16x4_t nw11_new = vdup_n_s16(iw11_new); const int row_stride = _input_new->info()->strides_in_bytes()[1]; const Coordinates top_left_window_corner_old(static_cast(old_keypoint_int_x) - _window_dimension / 2, static_cast(old_keypoint_int_y) - _window_dimension / 2); const Coordinates top_left_window_corner_new(static_cast(new_keypoint_int_x) - _window_dimension / 2, static_cast(new_keypoint_int_y) - _window_dimension / 2); const uint8_t *old_ptr = _input_old->buffer() + _input_old->info()->offset_element_in_bytes(top_left_window_corner_old); const uint8_t *new_ptr = _input_new->buffer() + _input_new->info()->offset_element_in_bytes(top_left_window_corner_new); static const int32x4_t nshifter_tensor = vdupq_n_s32(-(W_BITS - 5)); for(int ky = 0; ky < _window_dimension; ++ky, new_ptr += row_stride, old_ptr += row_stride) { int kx = 0; // Calculate elements in blocks of four as long as possible for(; kx <= _window_dimension - 4; kx += 4) { // Interpolation old tensor const int16x8_t nold_row1 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(old_ptr + kx))); const int16x8_t nold_row2 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(old_ptr + kx + row_stride))); const int32x4_t noldval = compute_bilinear_interpolation(nold_row1, nold_row2, nw00_old, nw01_old, nw10_old, nw11_old, nshifter_tensor); // Interpolation new tensor const int16x8_t nnew_row1 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(new_ptr + kx))); const int16x8_t nnew_row2 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(new_ptr + kx + row_stride))); const int32x4_t nnewval = compute_bilinear_interpolation(nnew_row1, nnew_row2, nw00_new, nw01_new, nw10_new, nw11_new, nshifter_tensor); // Calculate It gradient, i.e. pixelwise difference between old and new tensor const int32x4_t diff = vsubq_s32(nnewval, noldval); // Load the Ix and Iy gradient computed in the previous stage const int32x4_t nxval = vld1q_s32(bilinear_ix + kx + ky * _window_dimension); const int32x4_t nyval = vld1q_s32(bilinear_iy + kx + ky * _window_dimension); // Caculate Ix * It and Iy * It, and accumulate the results nb1 = vmlaq_s32(nb1, diff, nxval); nb2 = vmlaq_s32(nb2, diff, nyval); } // Calculate the leftover elements for(; kx < _window_dimension; ++kx) { const int32_t ival = get_pixel(_input_old, top_left_window_corner_old.x() + kx, top_left_window_corner_old.y() + ky, iw00_old, iw01_old, iw10_old, iw11_old, W_BITS - 5); const int32_t jval = get_pixel(_input_new, top_left_window_corner_new.x() + kx, top_left_window_corner_new.y() + ky, iw00_new, iw01_new, iw10_new, iw11_new, W_BITS - 5); const int32_t diff = jval - ival; ib1 += diff * bilinear_ix[kx + ky * _window_dimension]; ib2 += diff * bilinear_iy[kx + ky * _window_dimension]; } } ib1 += vgetq_lane_s32(nb1, 0) + vgetq_lane_s32(nb1, 1) + vgetq_lane_s32(nb1, 2) + vgetq_lane_s32(nb1, 3); ib2 += vgetq_lane_s32(nb2, 0) + vgetq_lane_s32(nb2, 1) + vgetq_lane_s32(nb2, 2) + vgetq_lane_s32(nb2, 3); return std::make_pair(ib1, ib2); } NELKTrackerKernel::NELKTrackerKernel() : _input_old(nullptr), _input_new(nullptr), _old_scharr_gx(nullptr), _old_scharr_gy(nullptr), _new_points(nullptr), _new_points_estimates(nullptr), _old_points(nullptr), _old_points_internal(), _new_points_internal(), _termination(Termination::TERM_CRITERIA_EPSILON), _use_initial_estimate(false), _pyramid_scale(0.0f), _epsilon(0.0f), _num_iterations(0), _window_dimension(0), _level(0), _num_levels(0), _valid_region() { } BorderSize NELKTrackerKernel::border_size() const { return BorderSize(1); } void NELKTrackerKernel::configure(const ITensor *input_old, const ITensor *input_new, const ITensor *old_scharr_gx, const ITensor *old_scharr_gy, const IKeyPointArray *old_points, const IKeyPointArray *new_points_estimates, IKeyPointArray *new_points, INELKInternalKeypointArray *old_points_internal, INELKInternalKeypointArray *new_points_internal, Termination termination, bool use_initial_estimate, float epsilon, unsigned int num_iterations, size_t window_dimension, size_t level, size_t num_levels, float pyramid_scale) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_old, 1, DataType::U8); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_new, 1, DataType::U8); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(old_scharr_gx, 1, DataType::S16); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(old_scharr_gy, 1, DataType::S16); _input_old = input_old; _input_new = input_new; _old_scharr_gx = old_scharr_gx; _old_scharr_gy = old_scharr_gy; _old_points = old_points; _new_points_estimates = new_points_estimates; _new_points = new_points; _old_points_internal = old_points_internal; _new_points_internal = new_points_internal; _termination = termination; _use_initial_estimate = use_initial_estimate; _epsilon = epsilon; _window_dimension = window_dimension; _level = level; _num_levels = num_levels; _pyramid_scale = pyramid_scale; _num_levels = num_levels; // Set maximum number of iterations used for convergence const size_t max_iterations = 1000; _num_iterations = (termination == Termination::TERM_CRITERIA_EPSILON) ? max_iterations : num_iterations; Window window; window.set(Window::DimX, Window::Dimension(0, old_points->num_values())); window.set(Window::DimY, Window::Dimension(0, 1)); _valid_region = intersect_valid_regions( input_old->info()->valid_region(), input_new->info()->valid_region(), old_scharr_gx->info()->valid_region(), old_scharr_gy->info()->valid_region()); update_window_and_padding(window, AccessWindowStatic(input_old->info(), _valid_region.start(0), _valid_region.start(1), _valid_region.end(0), _valid_region.end(1)), AccessWindowStatic(input_new->info(), _valid_region.start(0), _valid_region.start(1), _valid_region.end(0), _valid_region.end(1)), AccessWindowStatic(old_scharr_gx->info(), _valid_region.start(0), _valid_region.start(1), _valid_region.end(0), _valid_region.end(1)), AccessWindowStatic(old_scharr_gy->info(), _valid_region.start(0), _valid_region.start(1), _valid_region.end(0), _valid_region.end(1))); INEKernel::configure(window); } void NELKTrackerKernel::run(const Window &window, const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); ARM_COMPUTE_ERROR_ON(_input_old->buffer() == nullptr); ARM_COMPUTE_ERROR_ON(_input_new->buffer() == nullptr); ARM_COMPUTE_ERROR_ON(_old_scharr_gx->buffer() == nullptr); ARM_COMPUTE_ERROR_ON(_old_scharr_gy->buffer() == nullptr); const int list_end = window.x().end(); const int list_start = window.x().start(); init_keypoints(list_start, list_end); const int buffer_size = _window_dimension * _window_dimension; std::vector bilinear_ix(buffer_size); std::vector bilinear_iy(buffer_size); const int half_window = _window_dimension / 2; auto is_invalid_keypoint = [&](const NELKInternalKeypoint & keypoint) { const int x = std::floor(keypoint.x); const int y = std::floor(keypoint.y); return (x - half_window < _valid_region.start(0)) || (x + half_window >= _valid_region.end(0) - 1) || (y - half_window < _valid_region.start(1)) || (y + half_window >= _valid_region.end(1) - 1); }; for(int list_indx = list_start; list_indx < list_end; ++list_indx) { NELKInternalKeypoint &old_keypoint = _old_points_internal->at(list_indx); NELKInternalKeypoint &new_keypoint = _new_points_internal->at(list_indx); if(!old_keypoint.tracking_status) { continue; } if(is_invalid_keypoint(old_keypoint)) { if(_level == 0) { new_keypoint.tracking_status = false; } continue; } // Compute spatial gradient matrix int iA11 = 0; int iA12 = 0; int iA22 = 0; std::tie(iA11, iA12, iA22) = compute_spatial_gradient_matrix(old_keypoint, bilinear_ix.data(), bilinear_iy.data()); const float A11 = iA11 * FLT_SCALE; const float A12 = iA12 * FLT_SCALE; const float A22 = iA22 * FLT_SCALE; // Calculate minimum eigenvalue const float sum_A11_A22 = A11 + A22; const float discriminant = sum_A11_A22 * sum_A11_A22 - 4.0f * (A11 * A22 - A12 * A12); // Divide by _window_dimension^2 to reduce the floating point accummulation error const float minimum_eigenvalue = (sum_A11_A22 - std::sqrt(discriminant)) / (2.0f * _window_dimension * _window_dimension); // Determinant const double D = A11 * A22 - A12 * A12; // Check if it is a good point to track if(minimum_eigenvalue < EIGENVALUE_THRESHOLD || D < DETERMINANT_THRESHOLD) { // Invalidate tracked point if(_level == 0) { new_keypoint.tracking_status = false; } continue; } float prev_delta_x = 0.0f; float prev_delta_y = 0.0f; for(unsigned int j = 0; j < _num_iterations; ++j) { if(is_invalid_keypoint(new_keypoint)) { if(_level == 0) { new_keypoint.tracking_status = false; } break; } // Compute image mismatch vector int ib1 = 0; int ib2 = 0; std::tie(ib1, ib2) = compute_image_mismatch_vector(old_keypoint, new_keypoint, bilinear_ix.data(), bilinear_iy.data()); double b1 = ib1 * FLT_SCALE; double b2 = ib2 * FLT_SCALE; // Compute motion vector -> A^-1 * -b const float delta_x = (A12 * b2 - A22 * b1) / D; const float delta_y = (A12 * b1 - A11 * b2) / D; // Update the new position new_keypoint.x += delta_x; new_keypoint.y += delta_y; const float mag2 = delta_x * delta_x + delta_y * delta_y; // Check if termination criteria is EPSILON and if it is satisfied if(mag2 <= _epsilon && (_termination == Termination::TERM_CRITERIA_EPSILON || _termination == Termination::TERM_CRITERIA_BOTH)) { break; } // Check convergence analyzing the previous delta if(j > 0 && std::fabs(delta_x + prev_delta_x) < 0.01f && std::fabs(delta_y + prev_delta_y) < 0.01f) { new_keypoint.x -= delta_x * _pyramid_scale; new_keypoint.y -= delta_y * _pyramid_scale; break; } prev_delta_x = delta_x; prev_delta_y = delta_y; } } if(_level == 0) { for(int list_indx = list_start; list_indx < list_end; ++list_indx) { const NELKInternalKeypoint &new_keypoint = _new_points_internal->at(list_indx); _new_points->at(list_indx).x = roundf(new_keypoint.x); _new_points->at(list_indx).y = roundf(new_keypoint.y); _new_points->at(list_indx).tracking_status = new_keypoint.tracking_status ? 1 : 0; } } }