diff --git a/EyeTrackApp/blob.py b/EyeTrackApp/blob.py index 4bbd015..c17b028 100644 --- a/EyeTrackApp/blob.py +++ b/EyeTrackApp/blob.py @@ -1,3 +1,31 @@ +''' +------------------------------------------------------------------------------------------------------ + + ,@@@@@@ + @@@@@@@@@@@ @@@ + @@@@@@@@@@@@ @@@@@@@@@@@ + @@@@@@@@@@@@@ @@@@@@@@@@@@@@ + @@@@@@@/ ,@@@@@@@@@@@@@ + /@@@@@@@@@@@@@@@ @@@@@@@@ + @@@@@@@@@@@@@@@@@@@@@@@@ @@@@@ + @@@@@@@@ @@@@@ + ,@@@ @@@@& + @@@@@@. @@@@ + @@@ @@@@@@@@@/ @@@@@ + ,@@@. @@@@@@((@ @@@@( + //@@@ ,, @@@@ @@@@@ + @@@( @@@@@@@ + @@@ @ @@@@@@@@# + @@@@@@@@@@@@@@@@@ + @@@@@@@@@@@@@( + +BLOB By: Prohurtz#0001 (Main App Developer) +Algorithm App Implimentations By: Prohurtz#0001, qdot (Inital App Creator) + +Copyright (c) 2022 EyeTrackVR <3 +------------------------------------------------------------------------------------------------------ +''' + import cv2 import numpy as np diff --git a/EyeTrackApp/eye_processor.py b/EyeTrackApp/eye_processor.py index 469508f..e314a1e 100644 --- a/EyeTrackApp/eye_processor.py +++ b/EyeTrackApp/eye_processor.py @@ -53,12 +53,15 @@ if sys.platform.startswith("win"): from osc_calibrate_filter import * from haar_surround_feature import * from blob import * +from ransac import * +from hsrac import * class InformationOrigin(Enum): RANSAC = 1 BLOB = 2 FAILURE = 3 HSF = 4 + HSRAC = 5 bbb = 0 @dataclass @@ -91,111 +94,6 @@ async def delayed_setting_change(setting, value): -def ellipse_model(data, y, f): - """ - There is no need to make this process a function, since making the process a function will slow it down a little by calling it. - The results may be slightly different from the lambda version due to calculation errors derived from float types, but the calculation results are virtually the same. - a = 1.0,b = P[0],c = P[1],d = P[2],e = P[3],f = P[4] - :param data: - :param y: np.c_[d, e, a, c, b] - :param f: f == P[4, 0] - :return: this_return == np.array([ellipse_model(x, y) for (x, y) in data ]) - """ - return data.dot(y) + f - - -# @profile -def fit_rotated_ellipse_ransac(data: np.ndarray, rng: np.random.Generator, iter=100, sample_num=10, offset=80 # 80.0, 10, 80 - ): # before changing these values, please read up on the ransac algorithm - # However if you want to change any value just know that higher iterations will make processing frames slower - effective_sample = None - - # The array contents do not change during the loop, so only one call is needed. - # They say len is faster than shape. - # Reference url: https://stackoverflow.com/questions/35547853/what-is-faster-python3s-len-or-numpys-shape - len_data = len(data) - - if len_data < sample_num: - return None - - # Type of calculation result - ret_dtype = np.float64 - - # Sorts a random number array of size (iter,len_data). After sorting, returns the index of sample_num random numbers before sorting. - # If the array size is less than about 100, this is faster than rng.choice. - rng_sample = rng.random((iter, len_data)).argsort()[:, :sample_num] - # or - # I don't see any advantage to doing this. - # rng_sample = np.asarray(rng.random((iter, len_data)).argsort()[:, :sample_num], dtype=np.int32) - - # I don't think it looks beautiful. - # x,y,x**2,y**2,x*y,1,-1*x**2 - datamod = np.concatenate( - [data, data ** 2, (data[:, 0] * data[:, 1])[:, np.newaxis], np.ones((len_data, 1), dtype=ret_dtype), - (-1 * data[:, 0] ** 2)[:, np.newaxis]], axis=1, - dtype=ret_dtype) - - datamod_slim = np.array(datamod[:, :5], dtype=ret_dtype) - - datamod_rng = datamod[rng_sample] - datamod_rng6 = datamod_rng[:, :, 6] - datamod_rng_swap = datamod_rng[:, :, [4, 3, 0, 1, 5]] - datamod_rng_swap_trans = datamod_rng_swap.transpose((0, 2, 1)) - - # These two lines are one of the bottlenecks - datamod_rng_5x5 = np.matmul(datamod_rng_swap_trans, datamod_rng_swap) - datamod_rng_p5smp = np.matmul(np.linalg.inv(datamod_rng_5x5), datamod_rng_swap_trans) - - datamod_rng_p = np.matmul(datamod_rng_p5smp, datamod_rng6[:, :, np.newaxis]).reshape((-1, 5)) - - # I don't think it looks beautiful. - ellipse_y_arr = np.asarray( - [datamod_rng_p[:, 2], datamod_rng_p[:, 3], np.ones(len(datamod_rng_p)), datamod_rng_p[:, 1], datamod_rng_p[:, 0]], dtype=ret_dtype) - - ellipse_data_arr = ellipse_model(datamod_slim, ellipse_y_arr, np.asarray(datamod_rng_p[:, 4])).transpose((1, 0)) - ellipse_data_abs = np.abs(ellipse_data_arr) - ellipse_data_index = np.argmax(np.sum(ellipse_data_abs < offset, axis=1), axis=0) - effective_data_arr = ellipse_data_arr[ellipse_data_index] - effective_sample_p_arr = datamod_rng_p[ellipse_data_index] - - return fit_rotated_ellipse(effective_data_arr, effective_sample_p_arr) - - - -# @profile -def fit_rotated_ellipse(data, P): - a = 1.0 - b = P[0] - c = P[1] - d = P[2] - e = P[3] - f = P[4] - # The cost of trigonometric functions is high. - theta = 0.5 * np.arctan(b / (a - c), dtype=np.float64) - theta_sin = np.sin(theta, dtype=np.float64) - theta_cos = np.cos(theta, dtype=np.float64) - tc2 = theta_cos ** 2 - ts2 = theta_sin ** 2 - b_tcs = b * theta_cos * theta_sin - - # Do the calculation only once - cxy = b ** 2 - 4 * a * c - cx = (2 * c * d - b * e) / cxy - cy = (2 * a * e - b * d) / cxy - - # I just want to clear things up around here. - cu = a * cx ** 2 + b * cx * cy + c * cy ** 2 - f - cu_r = np.array([(a * tc2 + b_tcs + c * ts2), (a * ts2 - b_tcs + c * tc2)]) - wh = np.sqrt(cu / cu_r) - - w, h = wh[0], wh[1] - - error_sum = np.sum(data) - # print("fitting error = %.3f" % (error_sum)) - - return (cx, cy, w, h, theta) - - class EyeProcessor: def __init__( self, @@ -500,227 +398,6 @@ class EyeProcessor: - def HSRAC(self): - frame = self.current_image_gray - if self.now_mode == self.cv_mode[1]: - - - prev_res_len = len(self.response_list) - # adjustment of radius - if prev_res_len == 1: - # len==1==self.response_list==[self.settings.gui_HSF_radius] - self.cvparam.radius = self.auto_radius_range[0] - elif prev_res_len == 2: - # len==2==self.response_list==[self.settings.gui_HSF_radius, self.auto_radius_range[0]] - self.cvparam.radius = self.auto_radius_range[1] - elif prev_res_len == 3: - # len==3==self.response_list==[self.settings.gui_HSF_radius,self.auto_radius_range[0],self.auto_radius_range[1]] - sort_res = sorted(self.response_list, key=lambda x: x[1])[0] - # Extract the radius with the lowest response value - if sort_res[0] == self.settings.gui_HSF_radius: - # If the default value is best, change self.now_mode to init after setting radius to the default value. - self.cvparam.radius = self.settings.gui_HSF_radius - self.now_mode = self.cv_mode[2] if not self.skip_blink_detect else self.cv_mode[3] - self.response_list = [] - elif sort_res[0] == self.auto_radius_range[0]: - self.radius_cand_list = [i for i in range(self.auto_radius_range[0], self.settings.gui_HSF_radius, self.default_step[0])][1:] - # self.default_step is defined separately for xy, but radius is shared by xy, so it may be buggy - # It should be no problem to set it to anything other than self.default_step - self.cvparam.radius = self.radius_cand_list.pop() - else: - self.radius_cand_list = [i for i in range(self.settings.gui_HSF_radius, self.auto_radius_range[1], self.default_step[0])][1:] - # self.default_step is defined separately for xy, but radius is shared by xy, so it may be buggy - # It should be no problem to set it to anything other than self.default_step - self.cvparam.radius = self.radius_cand_list.pop() - else: - # Try the contents of the self.radius_cand_list in order until the self.radius_cand_list runs out - # Better make it a binary search. - if len(self.radius_cand_list) == 0: - sort_res = sorted(self.response_list, key=lambda x: x[1])[0] - self.cvparam.radius = sort_res[0] - self.now_mode = self.cv_mode[2] if not self.skip_blink_detect else self.cv_mode[3] - self.response_list = [] - else: - self.cvparam.radius = self.radius_cand_list.pop() - - radius, pad, step, hsf = self.cvparam.get_rpsh() - - # For measuring processing time of image processing - cv_start_time = timeit.default_timer() - - gray_frame = frame - - # Calculate the integral image of the frame - int_start_time = timeit.default_timer() - # BORDER_CONSTANT is faster than BORDER_REPLICATE There seems to be almost no negative impact when BORDER_CONSTANT is used. - frame_pad = cv2.copyMakeBorder(gray_frame, pad, pad, pad, pad, cv2.BORDER_CONSTANT) - frame_int = cv2.integral(frame_pad) - - # Convolve the feature with the integral image - conv_int_start_time = timeit.default_timer() - xy_step = frameint_get_xy_step(frame_int.shape, step, pad, start_offset=None, end_offset=None) - frame_conv, response, center_xy = conv_int(frame_int, hsf, step, pad, xy_step) - - crop_start_time = timeit.default_timer() - # Define the center point and radius - center_x, center_y = center_xy - upper_x = center_x + 25 #TODO make this a setting - lower_x = center_x - 25 - upper_y = center_y + 25 - lower_y = center_y - 25 - - # Crop the image using the calculated bounds - cropped_image = gray_frame[lower_y:upper_y, lower_x:upper_x] # y is 50px, x is 45? why? - - if self.now_mode == self.cv_mode[0] or self.now_mode == self.cv_mode[1]: - # If mode is first_frame or radius_adjust, record current radius and response - self.response_list.append((radius, response)) - elif self.now_mode == self.cv_mode[2]: - # Statistics for blink detection - if len(self.response_list) < self.blink_init_frames: - # Record the average value of cropped_image - self.response_list.append(cv2.mean(cropped_image)[0]) - else: - # Calculate self.response_max by computing interquartile range, IQR - # Change self.cv_mode to normal - self.response_list = np.array(self.response_list) - # 25%,75% - # This value may need to be adjusted depending on the environment. - quartile_1, quartile_3 = np.percentile(self.response_list, [25, 75]) - iqr = quartile_3 - quartile_1 - # response_min = quartile_1 - (iqr * 1.5) - self.response_max = quartile_3 + (iqr * 1.5) - self.now_mode = self.cv_mode[3] - else: - if 0 in cropped_image.shape: - # If shape contains 0, it is not detected well. - print("Something's wrong.") - else: - # If the average value of cropped_image is greater than self.response_max - # (i.e., if the cropimage is whitish - if self.response_max is not None and cv2.mean(cropped_image)[0] > self.response_max: - # blink - - cv2.circle(frame, (center_x, center_y), 10, (0, 0, 255), -1) - # If you want to update self.response_max. it may be more cost-effective to rewrite self.response_list in the following way - # https://stackoverflow.com/questions/42771110/fastest-way-to-left-cycle-a-numpy-array-like-pop-push-for-a-queue - - - #run ransac on the HSF crop\ - try: - kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3)) - thresh_add = 10 - rng = np.random.default_rng() - - f = False - - # Convert the image to grayscale, and set up thresholding. Thresholds here are basically a - # low-pass filter that will set any pixel < the threshold value to 0. Thresholding is user - # configurable in this utility as we're dealing with variable lighting amounts/placement, as - # well as camera positioning and lensing. Therefore everyone's cutoff may be different. - # - # The goal of thresholding settings is to make sure we can ONLY see the pupil. This is why we - # crop the image earlier; it gives us less possible dark area to get confused about in the - # next step. - frame = cropped_image - # For measuring processing time of image processing - # Crop first to reduce the amount of data to process. - frame = frame[0:len(frame) - 5, :] - # To reduce the processing data, first convert to 1-channel and then blur. - # The processing results were the same when I swapped the order of blurring and 1-channelization. - frame_gray = cv2.GaussianBlur(frame, (5, 5), 0) - - - # this will need to be adjusted everytime hardware is changed (brightness of IR, Camera postion, etc)m - min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(frame_gray) - - maxloc0_hf, maxloc1_hf = int(0.5 * max_loc[0]), int(0.5 * max_loc[1]) - - # crop 15% sqare around min_loc - # frame_gray = frame_gray[max_loc[1] - maxloc1_hf:max_loc[1] + maxloc1_hf, - # max_loc[0] - maxloc0_hf:max_loc[0] + maxloc0_hf] - - threshold_value = min_val + thresh_add - _, thresh = cv2.threshold(frame_gray, threshold_value, 255, cv2.THRESH_BINARY) - try: - opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel) - closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel) - th_frame = 255 - closing - except: - # I want to eliminate try here because try tends to be slow in execution. - th_frame = 255 - frame_gray - - - detect_start_time = timeit.default_timer() - contours, _ = cv2.findContours(th_frame, cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE) - hull = [] - # This way is faster than contours[i] - # But maybe this one is faster. hull = [cv2.convexHull(cnt, False) for cnt in contours] - for cnt in contours: - hull.append(cv2.convexHull(cnt, False)) - if not hull: - # If empty, go to next loop - pass - try: - - cnt = sorted(hull, key=cv2.contourArea) - maxcnt = cnt[-1] - # ellipse = cv2.fitEllipse(maxcnt) - ransac_data = fit_rotated_ellipse_ransac(maxcnt.reshape(-1, 2), rng) - if ransac_data is None: - # ransac_data is None==maxcnt.shape[0]') + + +class FPSResult(object): + """ + base https://github.com/ipython/ipython/blob/339c0d510a1f3cb2158dd8c6e7f4ac89aa4c89d8/IPython/core/magics/execution.py#L55 + """ + + def __init__(self, loops, repeat, best, worst, all_runs, precision): + self.loops = loops + self.repeat = repeat + self.best = 1 / best + self.worst = 1 / worst + self.all_runs = all_runs + self._precision = precision + self.fps = [1 / dt for dt in all_runs] + self.unit = "fps" + + @property + def average(self): + return math.fsum(self.fps) / len(self.fps) + + @property + def stdev(self): + mean = self.average + return (math.fsum([(x - mean) ** 2 for x in self.fps]) / len(self.fps)) ** 0.5 + + def __str__(self): + pm = '+-' + if hasattr(sys.stdout, 'encoding') and sys.stdout.encoding: + try: + u'\xb1'.encode(sys.stdout.encoding) + pm = u'\xb1' + except: + pass + return "min:{best} max:{worst} mean:{mean} {pm} {std} per loop (mean {pm} std. dev. of {runs} run{run_plural}, {loops:,} loop{loop_plural} each)".format( + pm=pm, + runs=self.repeat, + loops=self.loops, + loop_plural="" if self.loops == 1 else "s", + run_plural="" if self.repeat == 1 else "s", + mean="%.*g%s" % (self._precision, self.average, self.unit), + std="%.*g%s" % (self._precision, self.stdev, self.unit), + best="%.*g%s" % (self._precision, self.best, self.unit), + worst="%.*g%s" % (self._precision, self.worst, self.unit), + ) + + def _repr_pretty_(self, p, cycle): + unic = self.__str__() + p.text(u'') + + +def format_time(timespan, precision=3): + """ + https://github.com/ipython/ipython/blob/339c0d510a1f3cb2158dd8c6e7f4ac89aa4c89d8/IPython/core/magics/execution.py#L1473 + Formats the timespan in a human readable form + """ + + if timespan >= 60.0: + # we have more than a minute, format that in a human readable form + # Idea from http://snipplr.com/view/5713/ + parts = [("d", 60 * 60 * 24), ("h", 60 * 60), ("min", 60), ("s", 1)] + time = [] + leftover = timespan + for suffix, length in parts: + value = int(leftover / length) + if value > 0: + leftover = leftover % length + time.append(u'%s%s' % (str(value), suffix)) + if leftover < 1: + break + return " ".join(time) + + # Unfortunately the unicode 'micro' symbol can cause problems in + # certain terminals. + # See bug: https://bugs.launchpad.net/ipython/+bug/348466 + # Try to prevent crashes by being more secure than it needs to + # E.g. eclipse is able to print a ยต, but has no sys.stdout.encoding set. + units = [u"s", u"ms", u'us', "ns"] # the save value + if hasattr(sys.stdout, 'encoding') and sys.stdout.encoding: + try: + u'\xb5'.encode(sys.stdout.encoding) + units = [u"s", u"ms", u'\xb5s', "ns"] + except: + pass + scaling = [1, 1e3, 1e6, 1e9] + + if timespan > 0.0: + order = min(-int(math.floor(math.log10(timespan)) // 3), 3) + else: + order = 3 + return u"%.*g %s" % (precision, timespan * scaling[order], units[order]) + + +class CvParameters: + # It may be a little slower because a dict named "self" is read for each function call. + def __init__(self, radius, step): + # self.prev_radius=radius + self._radius = radius + self.pad = 2 * radius + # self.prev_step=step + self._step = step + self._hsf = HaarSurroundFeature(radius) + + def get_rpsh(self): + return self._radius, self.pad, self._step, self._hsf + # Essentially, the following would be preferable, but it would take twice as long to call. + # return self.radius, self.pad, self.step, self.hsf + + @property + def radius(self): + return self._radius + + @radius.setter + def radius(self, now_radius): + # self.prev_radius=self._radius + self._radius = now_radius + self.pad = 2 * now_radius + self.hsf = now_radius + + @property + def step(self): + return self._step + + @step.setter + def step(self, now_step): + # self.prev_step=self.step + self._step = now_step + + @property + def hsf(self): + return self._hsf + + @hsf.setter + def hsf(self, now_radius): + self._hsf = HaarSurroundFeature(now_radius) + + +class HaarSurroundFeature: + + def __init__(self, r_inner, r_outer=None, val=None): + if r_outer is None: + r_outer = r_inner * 3 + + r_inner2 = r_inner * r_inner + count_inner = r_inner2 + count_outer = r_outer * r_outer - r_inner2 + + if val is None: + val_inner = 1.0 / r_inner2 + val_outer = -val_inner * count_inner / count_outer + + else: + val_inner = val[0] + val_outer = val[1] + + self.val_in = np.array(val_inner, dtype=np.float64) + self.val_out = np.array(val_outer, dtype=np.float64) + self.r_in = r_inner + self.r_out = r_outer + + def get_kernel(self): + # Defined here, but not yet used? + # Create a kernel filled with the value of self.val_out + kernel = np.ones(shape=(2 * self.r_out - 1, 2 * self.r_out - 1), dtype=np.float64) * self.val_out + + # Set the values of the inner area of the kernel using array slicing + start = (self.r_out - self.r_in) + end = (self.r_out + self.r_in - 1) + kernel[start:end, start:end] = self.val_in + + return kernel + + +def to_gray(frame): + # Faster by quitting checking if the input image is already grayscale + # Perhaps it would be faster with less overhead to call cv2.cvtColor directly instead of using this function + return cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) + + +@lru_cache(maxsize=lru_maxsize_vs) +def frameint_get_xy_step(imageshape, xysteps, pad, start_offset=None, end_offset=None): + """ + + :param imageshape: (height(row),width(col)). row==y,cal==x + :param xysteps: (x,y) + :param pad: int + :param start_offset: (x,y) or None + :param end_offset: (x,y) or None + :return: xy_np:tuple(x,y) + """ + row, col = imageshape + row -= 1 + col -= 1 + x_step, y_step = xysteps + + # This is not beautiful. + start_pad_x = start_pad_y = end_pad_x = end_pad_y = pad + + if start_offset is not None: + start_pad_x += start_offset[0] + start_pad_y += start_offset[1] + if end_offset is not None: + end_pad_x += end_offset[0] + end_pad_y += end_offset[1] + y_np = np.arange(start_pad_y, row - end_pad_y, y_step) + x_np = np.arange(start_pad_x, col - end_pad_x, x_step) + + xy_np = (x_np, y_np) + + return xy_np + + +@lru_cache(maxsize=lru_maxsize_vvs) +def get_hsf_empty_array(len_syx, frameint_x, frame_int_dtype, fcshape): + # Function to reduce array allocation by providing an empty array first and recycling it with lru + inner_sum = np.empty(len_syx, dtype=frame_int_dtype) + outer_sum = np.empty(len_syx, dtype=frame_int_dtype) + p_temp = np.empty((len_syx[0], frameint_x), dtype=frame_int_dtype) + p00 = np.empty(len_syx, dtype=frame_int_dtype) + p11 = np.empty(len_syx, dtype=frame_int_dtype) + p01 = np.empty(len_syx, dtype=frame_int_dtype) + p10 = np.empty(len_syx, dtype=frame_int_dtype) + response_list = np.empty(len_syx, dtype=np.float64) + frame_conv = np.zeros(shape=fcshape[0], dtype=np.uint8) + frame_conv_stride = frame_conv[::fcshape[1], ::fcshape[2]] + return (inner_sum, outer_sum), p_temp, (p00, p11, p01, p10), response_list, (frame_conv, frame_conv_stride) + + +# @profile +def conv_int(frame_int, kernel, xy_step, padding, xy_steps_list): + """ + + :param frame_int: + :param kernel: hsf + :param step: (x,y) + :param padding: int + :return: + """ + row, col = frame_int.shape + row -= 1 + col -= 1 + x_step, y_step = xy_step + # padding2 = 2 * padding + f_shape = row - 2 * padding, col - 2 * padding + r_in = kernel.r_in + + len_sx, len_sy = len(xy_steps_list[0]), len(xy_steps_list[1]) + inout_sum, p_temp, p_list, response_list, frameconvlist = get_hsf_empty_array((len_sy, len_sx), col + 1, + frame_int.dtype, (f_shape, y_step, x_step)) + inner_sum, outer_sum = inout_sum + p00, p11, p01, p10 = p_list + frame_conv, frame_conv_stride = frameconvlist + + y_rin_m = xy_steps_list[1] - r_in + x_rin_m = xy_steps_list[0] - r_in + y_rin_p = xy_steps_list[1] + r_in + x_rin_p = xy_steps_list[0] + r_in + # xx==(y,x),m==MINUS,p==PLUS, ex: mm==(y-,x-) + inarr_mm = frame_int[y_rin_m[0]:y_rin_m[-1] + 1:y_step, x_rin_m[0]:x_rin_m[-1] + 1:x_step] + inarr_mp = frame_int[y_rin_m[0]:y_rin_m[-1] + 1:y_step, x_rin_p[0]:x_rin_p[-1] + 1:x_step] + inarr_pm = frame_int[y_rin_p[0]:y_rin_p[-1] + 1:y_step, x_rin_m[0]:x_rin_m[-1] + 1:x_step] + inarr_pp = frame_int[y_rin_p[0]:y_rin_p[-1] + 1:y_step, x_rin_p[0]:x_rin_p[-1] + 1:x_step] + + # == inarr_mm + inarr_pp - inarr_mp - inarr_pm + inner_sum[:, :] = inarr_mm + inner_sum += inarr_pp + inner_sum -= inarr_mp + inner_sum -= inarr_pm + + # Bottleneck here, I want to make it smarter. Someone do it. + # (y,x) + # p00=max(y_ro_m,0),max(x_ro_m,0) + # p11=min(y_ro_p,ylim),min(x_ro_p,xlim) + # p01=max(y_ro_m,0),min(x_ro_p,xlim) + # p10=min(y_ro_p,ylim),max(x_ro_m,0) + y_ro_m = xy_steps_list[1] - kernel.r_out + x_ro_m = xy_steps_list[0] - kernel.r_out + y_ro_p = xy_steps_list[1] + kernel.r_out + x_ro_p = xy_steps_list[0] + kernel.r_out + # p00 calc + np.take(frame_int, y_ro_m, axis=0, mode="clip", out=p_temp) + np.take(p_temp, x_ro_m, axis=1, mode="clip", out=p00) + # p01 calc + np.take(p_temp, x_ro_p, axis=1, mode="clip", out=p01) + # p11 calc + np.take(frame_int, y_ro_p, axis=0, mode="clip", out=p_temp) + np.take(p_temp, x_ro_p, axis=1, mode="clip", out=p11) + # p10 calc + np.take(p_temp, x_ro_m, axis=1, mode="clip", out=p10) + # the point is this + # p00=np.take(np.take(frame_int, y_ro_m, axis=0, mode="clip"), x_ro_m, axis=1, mode="clip") + # p11=np.take(np.take(frame_int, y_ro_p, axis=0, mode="clip"), x_ro_p, axis=1, mode="clip") + # p01=np.take(np.take(frame_int, y_ro_m, axis=0, mode="clip"), x_ro_p, axis=1, mode="clip") + # p10=np.take(np.take(frame_int, y_ro_p, axis=0, mode="clip"), x_ro_m, axis=1, mode="clip") + + outer_sum[:, :] = p00 + p11 - p01 - p10 - inner_sum + + np.multiply(kernel.val_in, inner_sum, dtype=np.float64, out=response_list) + response_list += kernel.val_out * outer_sum + + # min_response, max_val, min_loc, max_loc = cv2.minMaxLoc(self.response_list) + min_response, _, min_loc, _ = cv2.minMaxLoc(response_list) + + center = ((xy_steps_list[0][min_loc[0]] - padding), (xy_steps_list[1][min_loc[1]] - padding)) + + frame_conv_stride[:, :] = response_list + # or + # frame_conv_stride[:, :] = self.response_list.astype(np.uint8) + + return frame_conv, min_response, center + + + +#RANSAC \/ +def ellipse_model(data, y, f): + """ + There is no need to make this process a function, since making the process a function will slow it down a little by calling it. + The results may be slightly different from the lambda version due to calculation errors derived from float types, but the calculation results are virtually the same. + a = 1.0,b = P[0],c = P[1],d = P[2],e = P[3],f = P[4] + :param data: + :param y: np.c_[d, e, a, c, b] + :param f: f == P[4, 0] + :return: this_return == np.array([ellipse_model(x, y) for (x, y) in data ]) + """ + return data.dot(y) + f + +# @profile +def fit_rotated_ellipse_ransac(data: np.ndarray, rng: np.random.Generator, iter=100, sample_num=10, offset=80 # 80.0, 10, 80 + ): # before changing these values, please read up on the ransac algorithm + # However if you want to change any value just know that higher iterations will make processing frames slower + effective_sample = None + + # The array contents do not change during the loop, so only one call is needed. + # They say len is faster than shape. + # Reference url: https://stackoverflow.com/questions/35547853/what-is-faster-python3s-len-or-numpys-shape + len_data = len(data) + + if len_data < sample_num: + return None + + # Type of calculation result + ret_dtype = np.float64 + + # Sorts a random number array of size (iter,len_data). After sorting, returns the index of sample_num random numbers before sorting. + # If the array size is less than about 100, this is faster than rng.choice. + rng_sample = rng.random((iter, len_data)).argsort()[:, :sample_num] + # or + # I don't see any advantage to doing this. + # rng_sample = np.asarray(rng.random((iter, len_data)).argsort()[:, :sample_num], dtype=np.int32) + + # I don't think it looks beautiful. + # x,y,x**2,y**2,x*y,1,-1*x**2 + datamod = np.concatenate( + [data, data ** 2, (data[:, 0] * data[:, 1])[:, np.newaxis], np.ones((len_data, 1), dtype=ret_dtype), + (-1 * data[:, 0] ** 2)[:, np.newaxis]], axis=1, + dtype=ret_dtype) + + datamod_slim = np.array(datamod[:, :5], dtype=ret_dtype) + + datamod_rng = datamod[rng_sample] + datamod_rng6 = datamod_rng[:, :, 6] + datamod_rng_swap = datamod_rng[:, :, [4, 3, 0, 1, 5]] + datamod_rng_swap_trans = datamod_rng_swap.transpose((0, 2, 1)) + + # These two lines are one of the bottlenecks + datamod_rng_5x5 = np.matmul(datamod_rng_swap_trans, datamod_rng_swap) + datamod_rng_p5smp = np.matmul(np.linalg.inv(datamod_rng_5x5), datamod_rng_swap_trans) + + datamod_rng_p = np.matmul(datamod_rng_p5smp, datamod_rng6[:, :, np.newaxis]).reshape((-1, 5)) + + # I don't think it looks beautiful. + ellipse_y_arr = np.asarray( + [datamod_rng_p[:, 2], datamod_rng_p[:, 3], np.ones(len(datamod_rng_p)), datamod_rng_p[:, 1], datamod_rng_p[:, 0]], dtype=ret_dtype) + + ellipse_data_arr = ellipse_model(datamod_slim, ellipse_y_arr, np.asarray(datamod_rng_p[:, 4])).transpose((1, 0)) + ellipse_data_abs = np.abs(ellipse_data_arr) + ellipse_data_index = np.argmax(np.sum(ellipse_data_abs < offset, axis=1), axis=0) + effective_data_arr = ellipse_data_arr[ellipse_data_index] + effective_sample_p_arr = datamod_rng_p[ellipse_data_index] + + return fit_rotated_ellipse(effective_data_arr, effective_sample_p_arr) + + +# @profile +def fit_rotated_ellipse(data, P): + a = 1.0 + b = P[0] + c = P[1] + d = P[2] + e = P[3] + f = P[4] + # The cost of trigonometric functions is high. + theta = 0.5 * np.arctan(b / (a - c), dtype=np.float64) + theta_sin = np.sin(theta, dtype=np.float64) + theta_cos = np.cos(theta, dtype=np.float64) + tc2 = theta_cos ** 2 + ts2 = theta_sin ** 2 + b_tcs = b * theta_cos * theta_sin + + # Do the calculation only once + cxy = b ** 2 - 4 * a * c + cx = (2 * c * d - b * e) / cxy + cy = (2 * a * e - b * d) / cxy + + # I just want to clear things up around here. + cu = a * cx ** 2 + b * cx * cy + c * cy ** 2 - f + cu_r = np.array([(a * tc2 + b_tcs + c * ts2), (a * ts2 - b_tcs + c * tc2)]) + wh = np.sqrt(cu / cu_r) + + w, h = wh[0], wh[1] + + error_sum = np.sum(data) + # print("fitting error = %.3f" % (error_sum)) + + return (cx, cy, w, h, theta) + + + + +def HSRAC(self): + + frame = self.current_image_gray + if self.now_mode == self.cv_mode[1]: + + + prev_res_len = len(self.response_list) + # adjustment of radius + if prev_res_len == 1: + # len==1==self.response_list==[self.default_radius] + self.cvparam.radius = self.auto_radius_range[0] + elif prev_res_len == 2: + # len==2==self.response_list==[self.default_radius, self.auto_radius_range[0]] + self.cvparam.radius = self.auto_radius_range[1] + elif prev_res_len == 3: + # len==3==self.response_list==[self.default_radius,self.auto_radius_range[0],self.auto_radius_range[1]] + sort_res = sorted(self.response_list, key=lambda x: x[1])[0] + # Extract the radius with the lowest response value + if sort_res[0] == self.default_radius: + # If the default value is best, change self.now_mode to init after setting radius to the default value. + self.cvparam.radius = self.default_radius + self.now_mode = self.cv_mode[2] if not self.skip_blink_detect else self.cv_mode[3] + self.response_list = [] + elif sort_res[0] == self.auto_radius_range[0]: + self.radius_cand_list = [i for i in range(self.auto_radius_range[0], self.default_radius, self.default_step[0])][1:] + # self.default_step is defined separately for xy, but radius is shared by xy, so it may be buggy + # It should be no problem to set it to anything other than self.default_step + self.cvparam.radius = self.radius_cand_list.pop() + else: + self.radius_cand_list = [i for i in range(self.default_radius, self.auto_radius_range[1], self.default_step[0])][1:] + # self.default_step is defined separately for xy, but radius is shared by xy, so it may be buggy + # It should be no problem to set it to anything other than self.default_step + self.cvparam.radius = self.radius_cand_list.pop() + else: + # Try the contents of the self.radius_cand_list in order until the self.radius_cand_list runs out + # Better make it a binary search. + if len(self.radius_cand_list) == 0: + sort_res = sorted(self.response_list, key=lambda x: x[1])[0] + self.cvparam.radius = sort_res[0] + self.now_mode = self.cv_mode[2] if not self.skip_blink_detect else self.cv_mode[3] + self.response_list = [] + else: + self.cvparam.radius = self.radius_cand_list.pop() + + radius, pad, step, hsf = self.cvparam.get_rpsh() + + # For measuring processing time of image processing + cv_start_time = timeit.default_timer() + + gray_frame = frame + + # Calculate the integral image of the frame + int_start_time = timeit.default_timer() + # BORDER_CONSTANT is faster than BORDER_REPLICATE There seems to be almost no negative impact when BORDER_CONSTANT is used. + frame_pad = cv2.copyMakeBorder(gray_frame, pad, pad, pad, pad, cv2.BORDER_CONSTANT) + frame_int = cv2.integral(frame_pad) + + # Convolve the feature with the integral image + conv_int_start_time = timeit.default_timer() + xy_step = frameint_get_xy_step(frame_int.shape, step, pad, start_offset=None, end_offset=None) + frame_conv, response, center_xy = conv_int(frame_int, hsf, step, pad, xy_step) + + crop_start_time = timeit.default_timer() + # Define the center point and radius + center_x, center_y = center_xy + upper_x = center_x + 25 #TODO make this a setting + lower_x = center_x - 25 + upper_y = center_y + 25 + lower_y = center_y - 25 + + # Crop the image using the calculated bounds + cropped_image = gray_frame[lower_y:upper_y, lower_x:upper_x] # y is 50px, x is 45? why? + + if self.now_mode == self.cv_mode[0] or self.now_mode == self.cv_mode[1]: + # If mode is first_frame or radius_adjust, record current radius and response + self.response_list.append((radius, response)) + elif self.now_mode == self.cv_mode[2]: + # Statistics for blink detection + if len(self.response_list) < self.blink_init_frames: + # Record the average value of cropped_image + self.response_list.append(cv2.mean(cropped_image)[0]) + else: + # Calculate self.response_max by computing interquartile range, IQR + # Change self.cv_mode to normal + self.response_list = np.array(self.response_list) + # 25%,75% + # This value may need to be adjusted depending on the environment. + quartile_1, quartile_3 = np.percentile(self.response_list, [25, 75]) + iqr = quartile_3 - quartile_1 + # response_min = quartile_1 - (iqr * 1.5) + self.response_max = quartile_3 + (iqr * 1.5) + self.now_mode = self.cv_mode[3] + else: + if 0 in cropped_image.shape: + # If shape contains 0, it is not detected well. + print("Something's wrong.") + else: + # If the average value of cropped_image is greater than self.response_max + # (i.e., if the cropimage is whitish + if self.response_max is not None and cv2.mean(cropped_image)[0] > self.response_max: + # blink + + cv2.circle(frame, (center_x, center_y), 10, (0, 0, 255), -1) + # If you want to update self.response_max. it may be more cost-effective to rewrite self.response_list in the following way + # https://stackoverflow.com/questions/42771110/fastest-way-to-left-cycle-a-numpy-array-like-pop-push-for-a-queue + + +#run ransac on the HSF crop\ + try: + kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3)) + thresh_add = 10 + rng = np.random.default_rng() + + f = False + + # Convert the image to grayscale, and set up thresholding. Thresholds here are basically a + # low-pass filter that will set any pixel < the threshold value to 0. Thresholding is user + # configurable in this utility as we're dealing with variable lighting amounts/placement, as + # well as camera positioning and lensing. Therefore everyone's cutoff may be different. + # + # The goal of thresholding settings is to make sure we can ONLY see the pupil. This is why we + # crop the image earlier; it gives us less possible dark area to get confused about in the + # next step. + frame = cropped_image + # For measuring processing time of image processing + # Crop first to reduce the amount of data to process. + + #frame = frame[0:len(frame) - 5, :] + + # To reduce the processing data, first convert to 1-channel and then blur. + # The processing results were the same when I swapped the order of blurring and 1-channelization. + frame_gray = cv2.GaussianBlur(frame, (5, 5), 0) + + + # this will need to be adjusted everytime hardware is changed (brightness of IR, Camera postion, etc)m + min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(frame_gray) + + maxloc0_hf, maxloc1_hf = int(0.5 * max_loc[0]), int(0.5 * max_loc[1]) + + # crop 15% sqare around min_loc + # frame_gray = frame_gray[max_loc[1] - maxloc1_hf:max_loc[1] + maxloc1_hf, + # max_loc[0] - maxloc0_hf:max_loc[0] + maxloc0_hf] + + threshold_value = min_val + thresh_add + _, thresh = cv2.threshold(frame_gray, threshold_value, 255, cv2.THRESH_BINARY) + try: + opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel) + closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel) + th_frame = 255 - closing + except: + # I want to eliminate try here because try tends to be slow in execution. + th_frame = 255 - frame_gray + + + detect_start_time = timeit.default_timer() + contours, _ = cv2.findContours(th_frame, cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE) + hull = [] + # This way is faster than contours[i] + # But maybe this one is faster. hull = [cv2.convexHull(cnt, False) for cnt in contours] + for cnt in contours: + hull.append(cv2.convexHull(cnt, False)) + if not hull: + # If empty, go to next loop + pass + try: + + cnt = sorted(hull, key=cv2.contourArea) + maxcnt = cnt[-1] + # ellipse = cv2.fitEllipse(maxcnt) + ransac_data = fit_rotated_ellipse_ransac(maxcnt.reshape(-1, 2), rng) + if ransac_data is None: + # ransac_data is None==maxcnt.shape[0]