Experimental module for SnekStudio attempting to understand and potentially replace the built-in MediaPipe module
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 

594 lines
22 KiB

#!/usr/bin/python3
import copy
import gc
import json
import os
import re
import socket
import sys
import threading
import time
import traceback
import cv2
import mediapipe
import numpy
BaseOptions = mediapipe.tasks.BaseOptions
FaceLandmarker = mediapipe.tasks.vision.FaceLandmarker
FaceLandmarkerOptions = mediapipe.tasks.vision.FaceLandmarkerOptions
FaceLandmarkerResult = mediapipe.tasks.vision.FaceLandmarkerResult
# PoseLandmarker = mediapipe.tasks.vision.PoseLandmarker
# PoseLandmarkerOptions = mediapipe.tasks.vision.PoseLandmarkerOptions
HandLandmarker = mediapipe.tasks.vision.HandLandmarker
HandLandmarkerOptions = mediapipe.tasks.vision.HandLandmarkerOptions
HandLandmarkerResult = mediapipe.tasks.vision.HandLandmarkerResult
RunningMode = mediapipe.tasks.vision.RunningMode
# Indices of hand landmarks.
WRIST = 0
THUMB_CMC = 1
THUMB_MCP = 2
THUMB_IP = 3
THUMB_TIP = 4
INDEX_FINGER_MCP = 5
INDEX_FINGER_PIP = 6
INDEX_FINGER_DIP = 7
INDEX_FINGER_TIP = 8
MIDDLE_FINGER_MCP = 9
MIDDLE_FINGER_PIP = 10
MIDDLE_FINGER_DIP = 12
MIDDLE_FINGER_TIP = 13
RING_FINGER_MCP = 14
RING_FINGER_PIP = 15
RING_FINGER_DIP = 16
RING_FINGER_TIP = 17
PINKY_MCP = 18
PINKY_PIP = 19
PINKY_DIP = 20
PINKY_TIP = 21
DEFAULT_TRACKING_DATA = {
"face" : {
"confidence" : 0.0, # Currently either 0.0 or 1.0.
"transform" : [ [ 1.0, 0.0, 0.0, 0.0 ],
[ 0.0, 1.0, 0.0, 0.0 ],
[ 0.0, 0.0, 1.0, 0.0 ],
[ 0.0, 0.0, 0.0, 1.0 ], ],
"blendshapes" : {
"_neutral" : 0.0,
"browDownLeft" : 0.0,
"browDownRight" : 0.0,
"browInnerUp" : 0.0,
"browOuterUpLeft" : 0.0,
"browOuterUpRight" : 0.0,
"cheekPuff" : 0.0,
"cheekSquintLeft" : 0.0,
"cheekSquintRight" : 0.0,
"eyeBlinkLeft" : 0.0,
"eyeBlinkRight" : 0.0,
"eyeLookDownLeft" : 0.0,
"eyeLookDownRight" : 0.0,
"eyeLookInLeft" : 0.0,
"eyeLookInRight" : 0.0,
"eyeLookOutLeft" : 0.0,
"eyeLookOutRight" : 0.0,
"eyeLookUpLeft" : 0.0,
"eyeLookUpRight" : 0.0,
"eyeSquintLeft" : 0.0,
"eyeSquintRight" : 0.0,
"eyeWideLeft" : 0.0,
"eyeWideRight" : 0.0,
"jawForward" : 0.0,
"jawLeft" : 0.0,
"jawOpen" : 0.0,
"jawRight" : 0.0,
"mouthClose" : 0.0,
"mouthDimpleLeft" : 0.0,
"mouthDimpleRight" : 0.0,
"mouthFrownLeft" : 0.0,
"mouthFrownRight" : 0.0,
"mouthFunnel" : 0.0,
"mouthLeft" : 0.0,
"mouthLowerDownLeft" : 0.0,
"mouthLowerDownRight" : 0.0,
"mouthPressLeft" : 0.0,
"mouthPressRight" : 0.0,
"mouthPucker" : 0.0,
"mouthRight" : 0.0,
"mouthRollLower" : 0.0,
"mouthRollUpper" : 0.0,
"mouthShrugLower" : 0.0,
"mouthShrugUpper" : 0.0,
"mouthSmileLeft" : 0.0,
"mouthSmileRight" : 0.0,
"mouthStretchLeft" : 0.0,
"mouthStretchRight" : 0.0,
"mouthUpperUpLeft" : 0.0,
"mouthUpperUpRight" : 0.0,
"noseSneerLeft" : 0.0,
"noseSneerRight" : 0.0,
},
},
"hands" : {
"left" : {
"confidence" : 0.0,
"image_landmarks" : [ [ 0.0, 0.0, 0.0 ] ] * 21,
"world_landmarks" : [ [ 0.0, 0.0, 0.0 ] ] * 21,
},
"right" : {
"confidence" : 0.0,
"image_landmarks" : [ [ 0.0, 0.0, 0.0 ] ] * 21,
"world_landmarks" : [ [ 0.0, 0.0, 0.0 ] ] * 21,
},
},
}
class MediaPipeTracker:
def __init__(self):
self.the_big_ugly_mutex = threading.Lock()
self._tracker_worker_thread = None
# We need these to avoid deadlocks. If we're queueing frames
# faster than they can process, we'll hit a deadlock in
# MediaPipe.
self.frames_queued_face = 0
self.frames_queued_hands = 0
self.frames_queued_mutex = threading.Lock()
self.should_quit_threads = False
# Open the socket immediately so we can start sending error
# and status stuff to the hosting application.
self._udp_socket = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
self.udp_port_number = 7098
# FIXME: Make this editable.
self.minimum_frame_time = 0.016
self.video_device_index = -1
self.video_device_capture = None
self.landmarker = None
# self.landmarker_pose = None
self.landmarker_hands = None
# These are for more deadlock avoidance, so we can keep track
# of how behind the hand tracker is.
self._last_hand_result_timestamp = (time.time() * 1000)
self._last_hand_detect_timestamp = (time.time() * 1000)
self.output_data = copy.deepcopy(DEFAULT_TRACKING_DATA)
def _close_video_device(self):
with self.the_big_ugly_mutex:
self.video_device_capture = None
def _open_video_device(self):
with self.the_big_ugly_mutex:
if self.video_device_index == -1:
self.video_device_capture = None
return
# Check to make sure we don't already have the device open.
if self.video_device_capture != None:
return
# Try opening it!
self._write_log("Opening a video device!")
self.video_device_capture = cv2.VideoCapture(self.video_device_index)
# Enforce low-res capture for performance reasons.
try:
self.video_device_capture.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
self.video_device_capture.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
except Exception as e:
# Failed? Whatever. Just use the resolution it's stuck with.
pass
if self.video_device_capture.isOpened():
self._write_log("Video device acquired")
else:
self.video_device_capture = None
self._write_log("Failed to open video device: %s" % str(self.video_device_index))
def _init_mediapipe(self):
asset_path = os.path.abspath(os.path.dirname(__file__))
face_landmarker_path = os.path.join(asset_path, "face_landmarker.task")
# FIXME: Last minute breakages.
# pose_landmarker_path = os.path.join(asset_path, "pose_landmarker.task")
hand_landmarker_path = os.path.join(asset_path, "hand_landmarker.task")
options = FaceLandmarkerOptions(
base_options = BaseOptions(model_asset_path = face_landmarker_path),
running_mode = RunningMode.LIVE_STREAM,
output_face_blendshapes = True,
output_facial_transformation_matrixes = True,
result_callback = self._handle_result_face)
# FIXME: Last minute breakages.
# options_pose = PoseLandmarkerOptions(
# base_options = BaseOptions(model_asset_path = pose_landmarker_path),
# running_mode = RunningMode.LIVE_STREAM,
# output_segmentation_masks = False,
# result_callback = self._handle_result_pose)
options_hands = HandLandmarkerOptions(
base_options = BaseOptions(model_asset_path = hand_landmarker_path),
running_mode = RunningMode.LIVE_STREAM,
num_hands = 2,
# FIXME: Make these adjustable.
# Were working in the 4.1 version.
min_hand_detection_confidence = 0.75,
min_tracking_confidence = 0.75,
min_hand_presence_confidence = 0.9,
result_callback = self._handle_result_hands)
self._shutdown_mediapipe()
self._write_log("Init face landmarker...")
self.landmarker = FaceLandmarker.create_from_options(options)
# self._write_log("Init pose landmarker...")
# self.landmarker_pose = PoseLandmarker.create_from_options(options_pose)
self._write_log("Init hand landmarker...")
self.landmarker_hands = HandLandmarker.create_from_options(options_hands)
self._write_log("Init done")
def _write_log(self, *args):
try:
print(*args)
except Exception as e:
# Concerning...
pass
try:
self._send_status_packet(" ".join(str(s) for s in args))
except Exception as e:
pass
def _send_status_packet(self, status_str):
output_data = { "status" : status_str }
output_data_json = json.dumps(output_data, indent=4).encode("utf-8")
self._udp_socket.sendto(output_data_json, ("127.0.0.1", self.udp_port_number))
# Create a face landmarker instance with the live stream mode:
def _handle_result_face(
self,
result: FaceLandmarkerResult,
output_image: mediapipe.Image,
timestamp_ms: int,
):
with self.frames_queued_mutex:
self.frames_queued_face -= 1
face = self.output_data["face"]
face["confidence"] = 0.0
if len(result.facial_transformation_matrixes) > 0:
face["confidence"] = 1.0
face["transform"] = result.facial_transformation_matrixes[0].tolist()
if len(result.face_blendshapes) > 0:
face["confidence"] = 1.0
for shape in result.face_blendshapes[0]:
face["blendshapes"][shape.category_name] = shape.score
# FIXME: If we ever come back to it, finish this.
def _handle_result_pose(
self,
x,
output_image: mediapipe.Image,
timestamp_ms: int
):
for y in x.pose_world_landmarks:
pass
def _handle_result_hands(
self,
result: HandLandmarkerResult,
output_image: mediapipe.Image,
timestamp_ms: int,
):
with self.frames_queued_mutex:
self.frames_queued_hands -= 1
self._last_hand_result_timestamp = timestamp_ms
self.output_data["hands"]["left"]["confidence"] = 0.0
self.output_data["hands"]["right"]["confidence"] = 0.0
# TODO: Get actual dimensions of the camera?
frame_height, frame_width = (640, 480)
focal_length = frame_width * 0.75
center = (frame_width / 2, frame_height / 2)
camera_matrix = numpy.array([
[ focal_length, 0, center[0] ],
[ 0, focal_length, center[1] ],
[ 0, 0, 1 ]
], dtype = "double")
distortion = numpy.zeros((4, 1))
for index in range(len(result.hand_landmarks)):
handedness = result.handedness[index][0]
image_landmarks = result.hand_landmarks[index]
world_landmarks = result.hand_world_landmarks[index]
side = handedness.category_name.lower()
hand = self.output_data["hands"][side]
hand["confidence"] = handedness.score
for [i, image_landmark] in enumerate(image_landmarks):
world_landmark = world_landmarks[i]
hand["image_landmarks"][i] = [ image_landmark.x, image_landmark.y, image_landmark.z ]
hand["world_landmarks"][i] = [ world_landmark.x, world_landmark.y, world_landmark.z ]
def _tracker_worker_thread_func(self):
try:
# Deadlock-avoidance.
self._write_log("locking mutex before init mediapipe")
with self.the_big_ugly_mutex:
self._init_mediapipe()
self._write_log("Initializing MediaPipe")
self.output_data = copy.deepcopy(DEFAULT_TRACKING_DATA)
input_image = None
success = True
start_time = time.time()
frame_count = 0
# We'll send this when we're panicking from too many frames queued, as
# a last-ditch attempt to un-clog the queue before we get a deadlock
# thanks to the MediaPipe bug.
blank_image_cv2 = numpy.zeros((1,1,3), dtype=numpy.uint8)
blank_image_mp = mediapipe.Image(mediapipe.ImageFormat.SRGB, data=blank_image_cv2)
# Main capturing loop.
last_frame_time = 0
while not self.should_quit_threads:
# Wait for the minimum frame time.
time_to_sleep = self.minimum_frame_time - (time.time() - last_frame_time)
if time_to_sleep > 0.0:
time.sleep(time_to_sleep)
# If the video device got disconnected, reconnect it.
self._open_video_device()
with self.the_big_ugly_mutex:
last_frame_time = time.time()
last_timestamp_used = int(time.time() * 1000)
# Capture a frame.
if self.video_device_capture:
success, image = self.video_device_capture.read()
else:
# No camera connected at the moment. Just feed in
# blank images.
success = True
image = blank_image_cv2.copy()
if success:
# Convert image to MediaPipe.
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# FIXME: Find out why we do this. I think it was
# mentioned in the MediaPipe tutorial.
image.flags.writeable = False
mp_image = mediapipe.Image(
image_format=mediapipe.ImageFormat.SRGB,
data=image)
# Generate a timestamp to feed into the MediaPipe
# system. If we're still somehow inside the same
# millisecond as the last processed image, then skip
# this frame.
this_time = int(time.time() * 1000)
if this_time <= last_timestamp_used:
continue
# Check to see if we have too many face tracking
# frames queued.
need_reset = False
with self.frames_queued_mutex:
if self.frames_queued_face > 5:
need_reset = True
else:
self.frames_queued_face += 1
# Reset if we have too face frames queued. Avoid a
# deadlock.
if need_reset:
# Deadlock-avoidance.
self.landmarker._runner.restart()
self.frames_queued_face = 0
else:
self.landmarker.detect_async(mp_image, this_time)
# Hands
# If the last result we got back was too much time
# since the last one we queued up, then wait until
# some amount of time (which we guess in the most
# convoluted way possible) has passed.
#
# FIXME: Make this less stupid. Make it make
# sense. Then apply it to the face tracking.
hand_landmarker_time_skew = self._last_hand_detect_timestamp - self._last_hand_result_timestamp
if hand_landmarker_time_skew > 50: # FIXME: Make configurable (milliseconds)
self._last_hand_result_timestamp += this_time - self._last_hand_detect_timestamp
else:
# Check to see if we have too many hand tracking
# frames queued.
need_reset = False
with self.frames_queued_mutex:
if self.frames_queued_face > 5:
need_reset = True
else:
self.frames_queued_hands += 1
# If we do have too many frames queued, just reset
# the tracker to avoid a deadlock.
if need_reset:
self.landmarker_hands._runner.restart()
self.frames_queued_hands = 0
else:
self.landmarker_hands.detect_async(mp_image, this_time)
self._last_hand_detect_timestamp = this_time
# Track the last timestamp because we have to keep
# these monotonically increasing and we can't send
# the same timestamp twice.
last_timestamp_used = this_time
output_data_json = json.dumps(self.output_data, indent=4).encode("utf-8")
with self.frames_queued_mutex:
status_packet_str = "Tracking data sending. (Queue: %2d hand, %2d face)" % (self.frames_queued_hands, self.frames_queued_face)
self._write_log(status_packet_str)
# Output the packet.
self._udp_socket.sendto(output_data_json, ("127.0.0.1", self.udp_port_number))
self._write_log("Quitting")
except Exception as e:
exception_string_generator = traceback.TracebackException.from_exception(e)
exception_string = "".join(exception_string_generator.format())
self._write_log(exception_string)
def start_tracker(self):
if self._tracker_worker_thread:
stop_tracker()
assert(not self._tracker_worker_thread)
self._write_log("Starting worker thread.")
self._tracker_worker_thread = threading.Thread(
target=self._tracker_worker_thread_func,
daemon=True)
self._tracker_worker_thread.start()
self._write_log("Starting worker thread done.")
def stop_tracker(self):
assert(self._tracker_worker_thread)
self.should_quit_threads = True
self._write_log("Waiting for worker thread to join.")
self._tracker_worker_thread.join()
self._write_log("Worker thread joined.")
self._tracker_worker_thread = None
self.should_quit_threads = False
# Set to -1 to just release all devices.
def set_video_device_number(self, new_number):
if self.video_device_index != new_number:
with self.the_big_ugly_mutex:
self.video_device_index = new_number
self._close_video_device()
self._open_video_device()
def set_udp_port_number(self, new_number):
with self.the_big_ugly_mutex:
self.udp_port_number = new_number
def _shutdown_mediapipe(self):
if self.landmarker: self.landmarker.close()
# if self.landmarker_pose: self.landmarker_pose.close()
if self.landmarker_hands: self.landmarker_hands.close()
self.landmarker = None
# self.landmarker_pose = None
self.landmarker_hands = None
# Grumblegrumblegrumble...
gc.collect()
def __del__(self):
with self.the_big_ugly_mutex:
self._close_video_device()
self._shutdown_mediapipe()
# ----------------------------------------------------------------------
mediapipe_controller = MediaPipeTracker()
# ----------------------------------------------------------------------
# External interface (called from Godot)
def start_tracker():
global mediapipe_controller
mediapipe_controller.start_tracker()
def stop_tracker():
global mediapipe_controller
mediapipe_controller.stop_tracker()
# Set to -1 to just release all devices.
def set_video_device_number(new_number):
global mediapipe_controller
mediapipe_controller.set_video_device_number(new_number)
def set_udp_port_number(new_number):
global mediapipe_controller
mediapipe_controller.set_udp_port_number(new_number)
def enumerate_camera_devices():
from cv2_enumerate_cameras import enumerate_cameras
capture_api_preference=cv2.CAP_ANY
# Having issues with GSTREAMER sources, so let's just use V4L only.
if sys.platform == "linux": capture_api_preference = cv2.CAP_V4L2
# On Linux, we sometimes see stuff showing up as just "video#", so
# let's at least try to correlate paths and IDs from
# /dev/v4l/by-id .
path_to_name_mappings = {}
if sys.platform == "linux":
try:
device_id_list = os.listdir("/dev/v4l/by-id")
for device_id in device_id_list:
full_link_path = os.path.join("/dev/v4l/by-id", device_id)
actual_dev_file = os.path.abspath(os.path.join("/dev/v4l/by-id", os.readlink(full_link_path)))
path_to_name_mappings[actual_dev_file] = device_id
except IOError:
pass
all_camera_data = []
for camera_info in enumerate_cameras(apiPreference=capture_api_preference):
camera_name = camera_info.name
if re.match("video[0-9]+", camera_info.name):
if camera_info.path in path_to_name_mappings:
camera_name = path_to_name_mappings[camera_info.path]
# Figure out the backend.
backend_index = camera_info.backend
if sys.platform == "linux":
# For some reason, in Linux the backend is stored in the
# index and not the backend field.
backend_index = camera_info.index - (camera_info.index % 100)
backend_name = cv2.videoio_registry.getBackendName(backend_index)
camera_data = {
"name" : camera_name,
"backend" : backend_name,
"path" : camera_info.path,
"index" : camera_info.index,
}
all_camera_data.append(camera_data)
return all_camera_data