pose2sim/Pose2Sim/filtering.py
2024-10-04 13:39:28 +02:00

511 lines
18 KiB
Python

#!/usr/bin/env python
# -*- coding: utf-8 -*-
'''
###########################################################################
## FILTER 3D COORDINATES ##
###########################################################################
Filter trc 3D coordinates.
Available filters: Butterworth, Butterworth on speed, Gaussian, LOESS, Median
Set your parameters in Config.toml
INPUTS:
- a trc file
- filtering parameters in Config.toml
OUTPUT:
- a filtered trc file
'''
## INIT
import os
import glob
import fnmatch
import numpy as np
import pandas as pd
import cv2
import matplotlib.pyplot as plt
import logging
from scipy import signal
from scipy.ndimage import gaussian_filter1d
from statsmodels.nonparametric.smoothers_lowess import lowess
from filterpy.kalman import KalmanFilter, rts_smoother
from filterpy.common import Q_discrete_white_noise
from Pose2Sim.common import plotWindow
from Pose2Sim.common import convert_to_c3d
## AUTHORSHIP INFORMATION
__author__ = "David Pagnon"
__copyright__ = "Copyright 2021, Pose2Sim"
__credits__ = ["David Pagnon"]
__license__ = "BSD 3-Clause License"
__version__ = "0.9.4"
__maintainer__ = "David Pagnon"
__email__ = "contact@david-pagnon.com"
__status__ = "Development"
## FUNCTIONS
def kalman_filter(coords, frame_rate, measurement_noise, process_noise, nb_dimensions=3, nb_derivatives=3, smooth=True):
'''
Filters coordinates with a Kalman filter or a Kalman smoother
INPUTS:
- coords: array of shape (nframes, ndims)
- frame_rate: integer
- measurement_noise: integer
- process_noise: integer
- nb_dimensions: integer, number of dimensions (3 if 3D coordinates)
- nb_derivatives: integer, number of derivatives (3 if constant acceleration model)
- smooth: boolean. True if souble pass (recommended), False if single pass (if real-time)
OUTPUTS:
- kpt_coords_filt: filtered coords
'''
# Variables
dim_x = nb_dimensions * nb_derivatives # 9 state variables
dt = 1/frame_rate
# Filter definition
f = KalmanFilter(dim_x=dim_x, dim_z=nb_dimensions)
# States: initial position, velocity, accel, in 3D
def derivate_array(arr, dt=1):
return np.diff(arr, axis=0)/dt
def repeat(func, arg_func, nb_reps):
for i in range(nb_reps):
arg_func = func(arg_func)
return arg_func
x_init = []
for n_der in range(nb_derivatives):
x_init += [repeat(derivate_array, coords, n_der)[0]] # pose*3D, vel*3D, accel*3D
f.x = np.array(x_init).reshape(nb_dimensions,nb_derivatives).T.flatten() # pose, vel, accel *3D
# State transition matrix
F_per_coord = np.zeros((int(dim_x/nb_dimensions), int(dim_x/nb_dimensions)))
for i in range(nb_derivatives):
for j in range(min(i+1, nb_derivatives)):
F_per_coord[j,i] = dt**(i-j) / np.math.factorial(i - j)
f.F = np.kron(np.eye(nb_dimensions),F_per_coord)
# F_per_coord= [[1, dt, dt**2/2],
# [ 0, 1, dt ],
# [ 0, 0, 1 ]])
# No control input
f.B = None
# Measurement matrix (only positions)
H = np.zeros((nb_dimensions, dim_x))
for i in range(min(nb_dimensions,dim_x)):
H[i, int(i*(dim_x/nb_dimensions))] = 1
f.H = H
# H = [[1., 0., 0., 0., 0., 0., 0., 0., 0.],
# [0., 0., 0., 1., 0., 0., 0., 0., 0.],
# [0., 0., 0., 0., 0., 0., 1., 0., 0.]]
# Covariance matrix
f.P *= measurement_noise
# Measurement noise
f.R = np.diag([measurement_noise**2]*nb_dimensions)
# Process noise
f.Q = Q_discrete_white_noise(nb_derivatives, dt=dt, var=process_noise**2, block_size=nb_dimensions)
# Run filter: predict and update for each frame
mu, cov, _, _ = f.batch_filter(coords) # equivalent to below
# mu = []
# for kpt_coord_frame in coords:
# f.predict()
# f.update(kpt_coord_frame)
# mu.append(f.x.copy())
ind_of_position = [int(d*(dim_x/nb_dimensions)) for d in range(nb_dimensions)]
coords_filt = np.array(mu)[:,ind_of_position]
# RTS smoother
if smooth == True:
mu2, P, C, _ = f.rts_smoother(mu, cov)
coords_filt = np.array(mu2)[:,ind_of_position]
return coords_filt
def kalman_filter_1d(config_dict, frame_rate, col):
'''
1D Kalman filter
Deals with nans
INPUT:
- col: Pandas dataframe column
- trustratio: int, ratio process_noise/measurement_noise
- frame_rate: int
- smooth: boolean, True if double pass (recommended), False if single pass (if real-time)
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
trustratio = int(config_dict.get('filtering').get('kalman').get('trust_ratio'))
smooth = int(config_dict.get('filtering').get('kalman').get('smooth'))
measurement_noise = 20
process_noise = measurement_noise * trustratio
# split into sequences of not nans
col_filtered = col.copy()
mask = np.isnan(col_filtered) | col_filtered.eq(0)
falsemask_indices = np.where(~mask)[0]
gaps = np.where(np.diff(falsemask_indices) > 1)[0] + 1
idx_sequences = np.split(falsemask_indices, gaps)
if idx_sequences[0].size > 0:
idx_sequences_to_filter = [seq for seq in idx_sequences]
# Filter each of the selected sequences
for seq_f in idx_sequences_to_filter:
col_filtered[seq_f] = kalman_filter(col_filtered[seq_f], frame_rate, measurement_noise, process_noise, nb_dimensions=1, nb_derivatives=3, smooth=smooth).flatten()
return col_filtered
def butterworth_filter_1d(config_dict, frame_rate, col):
'''
1D Zero-phase Butterworth filter (dual pass)
Deals with nans
INPUT:
- col: numpy array
- order: int
- cutoff: int
- frame_rate: int
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
type = 'low' #config_dict.get('filtering').get('butterworth').get('type')
order = int(config_dict.get('filtering').get('butterworth').get('order'))
cutoff = int(config_dict.get('filtering').get('butterworth').get('cut_off_frequency'))
b, a = signal.butter(order/2, cutoff/(frame_rate/2), type, analog = False)
padlen = 3 * max(len(a), len(b))
# split into sequences of not nans
col_filtered = col.copy()
mask = np.isnan(col_filtered) | col_filtered.eq(0)
falsemask_indices = np.where(~mask)[0]
gaps = np.where(np.diff(falsemask_indices) > 1)[0] + 1
idx_sequences = np.split(falsemask_indices, gaps)
if idx_sequences[0].size > 0:
idx_sequences_to_filter = [seq for seq in idx_sequences if len(seq) > padlen]
# Filter each of the selected sequences
for seq_f in idx_sequences_to_filter:
col_filtered[seq_f] = signal.filtfilt(b, a, col_filtered[seq_f])
return col_filtered
def butterworth_on_speed_filter_1d(config_dict, frame_rate, col):
'''
1D zero-phase Butterworth filter (dual pass) on derivative
INPUT:
- col: Pandas dataframe column
- frame rate, order, cut-off frequency, type (from Config.toml)
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
type = 'low' # config_dict.get('filtering').get('butterworth_on_speed').get('type')
order = int(config_dict.get('filtering').get('butterworth_on_speed').get('order'))
cutoff = int(config_dict.get('filtering').get('butterworth_on_speed').get('cut_off_frequency'))
b, a = signal.butter(order/2, cutoff/(frame_rate/2), type, analog = False)
padlen = 3 * max(len(a), len(b))
# derivative
col_filtered = col.copy()
col_filtered_diff = col_filtered.diff() # derivative
col_filtered_diff = col_filtered_diff.fillna(col_filtered_diff.iloc[1]/2) # set first value correctly instead of nan
# split into sequences of not nans
mask = np.isnan(col_filtered_diff) | col_filtered_diff.eq(0)
falsemask_indices = np.where(~mask)[0]
gaps = np.where(np.diff(falsemask_indices) > 1)[0] + 1
idx_sequences = np.split(falsemask_indices, gaps)
if idx_sequences[0].size > 0:
idx_sequences_to_filter = [seq for seq in idx_sequences if len(seq) > padlen]
# Filter each of the selected sequences
for seq_f in idx_sequences_to_filter:
col_filtered_diff[seq_f] = signal.filtfilt(b, a, col_filtered_diff[seq_f])
col_filtered = col_filtered_diff.cumsum() + col.iloc[0] # integrate filtered derivative
return col_filtered
def gaussian_filter_1d(config_dict, frame_rate, col):
'''
1D Gaussian filter
INPUT:
- col: Pandas dataframe column
- gaussian_filter_sigma_kernel: kernel size from Config.toml
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
gaussian_filter_sigma_kernel = int(config_dict.get('filtering').get('gaussian').get('sigma_kernel'))
col_filtered = gaussian_filter1d(col, gaussian_filter_sigma_kernel)
return col_filtered
def loess_filter_1d(config_dict, frame_rate, col):
'''
1D LOWESS filter (Locally Weighted Scatterplot Smoothing)
INPUT:
- col: Pandas dataframe column
- loess_filter_nb_values: window used for smoothing from Config.toml
frac = loess_filter_nb_values * frames_number
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
kernel = config_dict.get('filtering').get('LOESS').get('nb_values_used')
col_filtered = col.copy()
mask = np.isnan(col_filtered)
falsemask_indices = np.where(~mask)[0]
gaps = np.where(np.diff(falsemask_indices) > 1)[0] + 1
idx_sequences = np.split(falsemask_indices, gaps)
if idx_sequences[0].size > 0:
idx_sequences_to_filter = [seq for seq in idx_sequences if len(seq) > kernel]
# Filter each of the selected sequences
for seq_f in idx_sequences_to_filter:
col_filtered[seq_f] = lowess(col_filtered[seq_f], seq_f, is_sorted=True, frac=kernel/len(seq_f), it=0)[:,1]
return col_filtered
def median_filter_1d(config_dict, frame_rate, col):
'''
1D median filter
INPUT:
- col: Pandas dataframe column
- median_filter_kernel_size: kernel size from Config.toml
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
median_filter_kernel_size = config_dict.get('filtering').get('median').get('kernel_size')
col_filtered = signal.medfilt(col, kernel_size=median_filter_kernel_size)
return col_filtered
def display_figures_fun(Q_unfilt, Q_filt, time_col, keypoints_names, person_id=0):
'''
Displays filtered and unfiltered data for comparison
INPUTS:
- Q_unfilt: pandas dataframe of unfiltered 3D coordinates
- Q_filt: pandas dataframe of filtered 3D coordinates
- time_col: pandas column
- keypoints_names: list of strings
OUTPUT:
- matplotlib window with tabbed figures for each keypoint
'''
pw = plotWindow()
pw.MainWindow.setWindowTitle('Person '+ str(person_id) + ' coordinates')
for id, keypoint in enumerate(keypoints_names):
f = plt.figure()
axX = plt.subplot(311)
plt.plot(time_col.to_numpy(), Q_unfilt.iloc[:,id*3].to_numpy(), label='unfiltered')
plt.plot(time_col.to_numpy(), Q_filt.iloc[:,id*3].to_numpy(), label='filtered')
plt.setp(axX.get_xticklabels(), visible=False)
axX.set_ylabel(keypoint+' X')
plt.legend()
axY = plt.subplot(312)
plt.plot(time_col.to_numpy(), Q_unfilt.iloc[:,id*3+1].to_numpy(), label='unfiltered')
plt.plot(time_col.to_numpy(), Q_filt.iloc[:,id*3+1].to_numpy(), label='filtered')
plt.setp(axY.get_xticklabels(), visible=False)
axY.set_ylabel(keypoint+' Y')
plt.legend()
axZ = plt.subplot(313)
plt.plot(time_col.to_numpy(), Q_unfilt.iloc[:,id*3+2].to_numpy(), label='unfiltered')
plt.plot(time_col.to_numpy(), Q_filt.iloc[:,id*3+2].to_numpy(), label='filtered')
axZ.set_ylabel(keypoint+' Z')
axZ.set_xlabel('Time')
plt.legend()
pw.addPlot(keypoint, f)
pw.show()
def filter1d(col, config_dict, filter_type, frame_rate):
'''
Choose filter type and filter column
INPUT:
- col: Pandas dataframe column
- filter_type: filter type from Config.toml
- frame_rate: int
OUTPUT:
- col_filtered: Filtered pandas dataframe column
'''
# Choose filter
filter_mapping = {
'kalman': kalman_filter_1d,
'butterworth': butterworth_filter_1d,
'butterworth_on_speed': butterworth_on_speed_filter_1d,
'gaussian': gaussian_filter_1d,
'LOESS': loess_filter_1d,
'median': median_filter_1d
}
filter_fun = filter_mapping[filter_type]
# Filter column
col_filtered = filter_fun(config_dict, frame_rate, col)
return col_filtered
def recap_filter3d(config_dict, trc_path):
'''
Print a log message giving filtering parameters. Also stored in User/logs.txt.
OUTPUT:
- Message in console
'''
# Read Config
filter_type = config_dict.get('filtering').get('type')
kalman_filter_trustratio = int(config_dict.get('filtering').get('kalman').get('trust_ratio'))
kalman_filter_smooth = int(config_dict.get('filtering').get('kalman').get('smooth'))
kalman_filter_smooth_str = 'smoother' if kalman_filter_smooth else 'filter'
butterworth_filter_type = 'low' # config_dict.get('filtering').get('butterworth').get('type')
butterworth_filter_order = int(config_dict.get('filtering').get('butterworth').get('order'))
butterworth_filter_cutoff = int(config_dict.get('filtering').get('butterworth').get('cut_off_frequency'))
butter_speed_filter_type = 'low' # config_dict.get('filtering').get('butterworth_on_speed').get('type')
butter_speed_filter_order = int(config_dict.get('filtering').get('butterworth_on_speed').get('order'))
butter_speed_filter_cutoff = int(config_dict.get('filtering').get('butterworth_on_speed').get('cut_off_frequency'))
gaussian_filter_sigma_kernel = int(config_dict.get('filtering').get('gaussian').get('sigma_kernel'))
loess_filter_nb_values = config_dict.get('filtering').get('LOESS').get('nb_values_used')
median_filter_kernel_size = config_dict.get('filtering').get('median').get('kernel_size')
make_c3d = config_dict.get('filtering').get('make_c3d')
# Recap
filter_mapping_recap = {
'kalman': f'--> Filter type: Kalman {kalman_filter_smooth_str}. Measurements trusted {kalman_filter_trustratio} times as much as previous data, assuming a constant acceleration process.',
'butterworth': f'--> Filter type: Butterworth {butterworth_filter_type}-pass. Order {butterworth_filter_order}, Cut-off frequency {butterworth_filter_cutoff} Hz.',
'butterworth_on_speed': f'--> Filter type: Butterworth on speed {butter_speed_filter_type}-pass. Order {butter_speed_filter_order}, Cut-off frequency {butter_speed_filter_cutoff} Hz.',
'gaussian': f'--> Filter type: Gaussian. Standard deviation kernel: {gaussian_filter_sigma_kernel}',
'LOESS': f'--> Filter type: LOESS. Number of values used: {loess_filter_nb_values}',
'median': f'--> Filter type: Median. Kernel size: {median_filter_kernel_size}'
}
logging.info(filter_mapping_recap[filter_type])
logging.info(f'Filtered 3D coordinates are stored at {trc_path}.\n')
if make_c3d:
logging.info('All filtered trc files have been converted to c3d.')
def filter_all(config_dict):
'''
Filter the 3D coordinates of the trc file.
Displays filtered coordinates for checking.
INPUTS:
- a trc file
- filtration parameters from Config.toml
OUTPUT:
- a filtered trc file
'''
# Read config_dict
project_dir = config_dict.get('project').get('project_dir')
pose3d_dir = os.path.realpath(os.path.join(project_dir, 'pose-3d'))
display_figures = config_dict.get('filtering').get('display_figures')
filter_type = config_dict.get('filtering').get('type')
make_c3d = config_dict.get('filtering').get('make_c3d')
# Get frame_rate
video_dir = os.path.join(project_dir, 'videos')
vid_img_extension = config_dict['pose']['vid_img_extension']
video_files = glob.glob(os.path.join(video_dir, '*'+vid_img_extension))
frame_rate = config_dict.get('project').get('frame_rate')
if frame_rate == 'auto':
try:
cap = cv2.VideoCapture(video_files[0])
cap.read()
if cap.read()[0] == False:
raise
frame_rate = int(cap.get(cv2.CAP_PROP_FPS))
except:
frame_rate = 60
# Trc paths
trc_path_in = [file for file in glob.glob(os.path.join(pose3d_dir, '*.trc')) if 'filt' not in file]
trc_f_out = [f'{os.path.basename(t).split(".")[0]}_filt_{filter_type}.trc' for t in trc_path_in]
trc_path_out = [os.path.join(pose3d_dir, t) for t in trc_f_out]
for person_id, (t_in, t_out) in enumerate(zip(trc_path_in, trc_path_out)):
# Read trc header
with open(t_in, 'r') as trc_file:
header = [next(trc_file) for line in range(5)]
# Read trc coordinates values
trc_df = pd.read_csv(t_in, sep="\t", skiprows=4)
frames_col, time_col = trc_df.iloc[:,0], trc_df.iloc[:,1]
Q_coord = trc_df.drop(trc_df.columns[[0, 1, -1]], axis=1)
# Filter coordinates
Q_filt = Q_coord.apply(filter1d, axis=0, args = [config_dict, filter_type, frame_rate])
# Display figures
if display_figures:
# Retrieve keypoints
keypoints_names = pd.read_csv(t_in, sep="\t", skiprows=3, nrows=0).columns[2::3][:-1].to_numpy()
display_figures_fun(Q_coord, Q_filt, time_col, keypoints_names, person_id)
# Reconstruct trc file with filtered coordinates
with open(t_out, 'w') as trc_o:
[trc_o.write(line) for line in header]
Q_filt.insert(0, 'Frame#', frames_col)
Q_filt.insert(1, 'Time', time_col)
# Q_filt = Q_filt.fillna(' ')
Q_filt.to_csv(trc_o, sep='\t', index=False, header=None, lineterminator='\n')
# Save c3d
if make_c3d:
convert_to_c3d(t_out)
# Recap
recap_filter3d(config_dict, t_out)