robocar-oak-camera/camera/oak_pipeline.py

663 lines
26 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

"""
Camera event loop
"""
import abc
import datetime
import logging
import pathlib
import time
from dataclasses import dataclass
from typing import List, Any
import cv2
import depthai as dai
import events.events_pb2 as evt
import numpy as np
import numpy.typing as npt
import paho.mqtt.client as mqtt
from depthai import Device
logger = logging.getLogger(__name__)
_NN_PATH = "/models/mobile_object_localizer_192x192_openvino_2021.4_6shave.blob"
_NN_WIDTH = 192
_NN_HEIGHT = 192
_PREVIEW_WIDTH = 640
_PREVIEW_HEIGHT = 480
_CAMERA_BASELINE_IN_MM = 75
class ObjectProcessor:
"""
Processor for Object detection
"""
def __init__(self, mqtt_client: mqtt.Client, objects_topic: str, objects_threshold: float):
self._mqtt_client = mqtt_client
self._objects_topic = objects_topic
self._objects_threshold = objects_threshold
def process(self, in_nn: dai.NNData, frame_ref: evt.FrameRef) -> None:
"""
Parse and publish result of NeuralNetwork result
:param in_nn: NeuralNetwork result read from device
:param frame_ref: Id of the frame where objects are been detected
:return:
"""
detection_boxes = np.array(in_nn.getLayerFp16("ExpandDims")).reshape((100, 4))
detection_scores = np.array(in_nn.getLayerFp16("ExpandDims_2")).reshape((100,))
# keep boxes bigger than threshold
mask = detection_scores >= self._objects_threshold
boxes = detection_boxes[mask]
scores = detection_scores[mask]
if boxes.shape[0] > 0:
self._publish_objects(boxes, frame_ref, scores)
def _publish_objects(self, boxes: npt.NDArray[np.float64], frame_ref: evt.FrameRef, scores: npt.NDArray[np.float64]) -> None:
objects_msg = evt.ObjectsMessage()
objs = []
for i in range(boxes.shape[0]):
logger.debug("new object detected: %s", str(boxes[i]))
objs.append(_bbox_to_object(boxes[i], scores[i].astype(float)))
objects_msg.objects.extend(objs)
objects_msg.frame_ref.name = frame_ref.name
objects_msg.frame_ref.id = frame_ref.id
objects_msg.frame_ref.created_at.FromDatetime(frame_ref.created_at.ToDatetime())
logger.debug("publish object event to %s", self._objects_topic)
self._mqtt_client.publish(topic=self._objects_topic,
payload=objects_msg.SerializeToString(),
qos=0,
retain=False)
class FrameProcessError(Exception):
"""
Error base for invalid frame processing
Attributes:
message -- explanation of the error
"""
def __init__(self, message: str):
"""
:param message: explanation of the error
"""
self.message = message
class FrameProcessor:
"""
Processor for camera frames
"""
def __init__(self, mqtt_client: mqtt.Client, frame_topic: str):
self._mqtt_client = mqtt_client
self._frame_topic = frame_topic
def process(self, img: dai.ImgFrame) -> Any:
"""
Publish camera frames
:param img: image read from camera
:return:
id frame reference
:raise:
FrameProcessError if frame can't be processed
"""
im_resize = img.getCvFrame()
is_success, im_buf_arr = cv2.imencode(".jpg", im_resize)
if not is_success:
raise FrameProcessError("unable to process to encode frame to jpg")
byte_im = im_buf_arr.tobytes()
now = datetime.datetime.now()
frame_msg = evt.FrameMessage()
frame_msg.id.name = "robocar-oak-camera-oak"
frame_msg.id.id = str(int(now.timestamp() * 1000))
frame_msg.id.created_at.FromDatetime(now)
frame_msg.frame = byte_im
logger.debug("publish frame event to %s", self._frame_topic)
self._mqtt_client.publish(topic=self._frame_topic,
payload=frame_msg.SerializeToString(),
qos=0,
retain=False)
return frame_msg.id
class DisparityProcessor:
"""
Processor for camera frames
"""
def __init__(self, mqtt_client: mqtt.Client, disparity_topic: str):
self._mqtt_client = mqtt_client
self._disparity_topic = disparity_topic
def process(self, img: dai.ImgFrame, frame_ref: evt.FrameRef, focal_length_in_pixels: float,
baseline_mm: float = _CAMERA_BASELINE_IN_MM) -> None:
im_frame = img.getCvFrame()
is_success, im_buf_arr = cv2.imencode(".jpg", im_frame)
if not is_success:
raise FrameProcessError("unable to process to encode frame to jpg")
byte_im = im_buf_arr.tobytes()
disparity_msg = evt.DisparityMessage()
disparity_msg.disparity = byte_im
disparity_msg.frame_ref.name = frame_ref.name
disparity_msg.frame_ref.id = frame_ref.id
disparity_msg.frame_ref.created_at.FromDatetime(frame_ref.created_at.ToDatetime())
disparity_msg.focal_length_in_pixels = focal_length_in_pixels
disparity_msg.baseline_in_mm = baseline_mm
self._mqtt_client.publish(topic=self._disparity_topic,
payload=disparity_msg.SerializeToString(),
qos=0,
retain=False)
class Source(abc.ABC):
"""Base class for image source"""
@abc.abstractmethod
def get_stream_name(self) -> str:
"""
Queue/stream name to use to get data
:return: steam name
"""
@abc.abstractmethod
def link(self, input_node: dai.Node.Input) -> None:
"""
Link this source to the input node
:param: input_node: input node to link
"""
class ObjectDetectionNN:
"""
Node to detect objects into image
Read image as input and apply resize transformation before to run NN on it
Result is available with 'get_stream_name()' stream
"""
def __init__(self, pipeline: dai.Pipeline):
# Define a neural network that will make predictions based on the source frames
detection_nn = pipeline.createNeuralNetwork()
detection_nn.setBlobPath(pathlib.Path(_NN_PATH))
detection_nn.setNumPoolFrames(4)
detection_nn.input.setBlocking(False)
detection_nn.setNumInferenceThreads(2)
self._detection_nn = detection_nn
self._xout = self._configure_xout_nn(pipeline)
self._detection_nn.out.link(self._xout.input)
self._manip_image = self._configure_manip(pipeline)
self._manip_image.out.link(self._detection_nn.input)
@staticmethod
def _configure_manip(pipeline: dai.Pipeline) -> dai.node.ImageManip:
# Resize image
manip = pipeline.createImageManip()
manip.initialConfig.setResize(_NN_WIDTH, _NN_HEIGHT)
manip.initialConfig.setFrameType(dai.ImgFrame.Type.RGB888p)
manip.initialConfig.setKeepAspectRatio(False)
return manip
@staticmethod
def _configure_xout_nn(pipeline: dai.Pipeline) -> dai.node.XLinkOut:
xout_nn = pipeline.createXLinkOut()
xout_nn.setStreamName("nn")
xout_nn.input.setBlocking(False)
return xout_nn
def get_stream_name(self) -> str:
"""
Queue/stream name to use to get data
:return: stream name
"""
return self._xout.getStreamName()
def get_input(self) -> dai.Node.Input:
"""
Get input node to use to link with source node
:return: input to link with source output, see Source.link()
"""
return self._manip_image.inputImage
class CameraSource(Source):
"""Image source based on camera preview"""
def __init__(self, pipeline: dai.Pipeline, img_width: int, img_height: int, fps: int):
self._cam_rgb = pipeline.createColorCamera()
self._xout_rgb = pipeline.createXLinkOut()
self._xout_rgb.setStreamName("rgb")
# Properties
self._cam_rgb.setBoardSocket(dai.CameraBoardSocket.RGB)
self._cam_rgb.setPreviewSize(width=_PREVIEW_WIDTH, height=_PREVIEW_HEIGHT)
self._cam_rgb.setInterleaved(False)
self._cam_rgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.RGB)
self._cam_rgb.setFps(fps)
self._resize_manip = self._configure_manip(pipeline=pipeline, img_width=img_width, img_height=img_height)
# link camera preview to output
self._cam_rgb.preview.link(self._resize_manip.inputImage)
self._resize_manip.out.link(self._xout_rgb.input)
def link(self, input_node: dai.Node.Input) -> None:
self._cam_rgb.preview.link(input_node)
def get_stream_name(self) -> str:
return self._xout_rgb.getStreamName()
@staticmethod
def _configure_manip(pipeline: dai.Pipeline, img_width: int, img_height: int) -> dai.node.ImageManip:
# Resize image
manip = pipeline.createImageManip()
manip.initialConfig.setResize(img_width, img_height)
manip.initialConfig.setFrameType(dai.ImgFrame.Type.RGB888p)
manip.initialConfig.setKeepAspectRatio(False)
return manip
class StereoDepthPostFilter(abc.ABC):
@abc.abstractmethod
def apply(self, config: dai.RawStereoDepthConfig) -> None:
pass
class MedianFilter(StereoDepthPostFilter):
"""
This is a non-edge preserving Median filter, which can be used to reduce noise and smoothen the depth map.
Median filter is implemented in hardware, so its the fastest filter.
"""
def __init__(self, value: dai.MedianFilter = dai.MedianFilter.KERNEL_7x7) -> None:
self._value = value
def apply(self, config: dai.RawStereoDepthConfig) -> None:
config.postProcessing.median.value = self._value
class SpeckleFilter(StereoDepthPostFilter):
"""
Speckle Filter is used to reduce the speckle noise. Speckle noise is a region with huge variance between
neighboring disparity/depth pixels, and speckle filter tries to filter this region.
"""
def __init__(self, enable: bool = True, speckle_range: int = 50) -> None:
"""
:param enable: Whether to enable or disable the filter.
:param speckle_range: Speckle search range.
"""
self._enable = enable
self._speckle_range = speckle_range
def apply(self, config: dai.RawStereoDepthConfig) -> None:
config.postProcessing.speckleFilter.enable = self._enable
config.postProcessing.speckleFilter.speckleRange = self._speckle_range
class TemporalFilter(StereoDepthPostFilter):
"""
Temporal Filter is intended to improve the depth data persistency by manipulating per-pixel values based on
previous frames. The filter performs a single pass on the data, adjusting the depth values while also updating the
tracking history. In cases where the pixel data is missing or invalid, the filter uses a user-defined persistency
mode to decide whether the missing value should be rectified with stored data. Note that due to its reliance on
historic data the filter may introduce visible blurring/smearing artifacts, and therefore is best-suited for
static scenes.
"""
def __init__(self,
enable: bool = True,
persistencyMode: dai.RawStereoDepthConfig.PostProcessing.TemporalFilter.PersistencyMode=dai.RawStereoDepthConfig.PostProcessing.TemporalFilter.PersistencyMode.VALID_2_IN_LAST_4,
alpha: float = 0.4,
delta: int = 0):
"""
:param enable: Whether to enable or disable the filter.
:param persistencyMode: Persistency mode. If the current disparity/depth value is invalid, it will be replaced
by an older value, based on persistency mode.
:param alpha: The Alpha factor in an exponential moving average with Alpha=1 - no filter.
Alpha = 0 - infinite filter. Determines the extent of the temporal history that should be averaged.
:param delta: Step-size boundary. Establishes the threshold used to preserve surfaces (edges).
If the disparity value between neighboring pixels exceed the disparity threshold set by this delta parameter,
then filtering will be temporarily disabled. Default value 0 means auto: 3 disparity integer levels.
In case of subpixel mode its 3*number of subpixel levels.
"""
self._enable = enable
self._persistencyMode = persistencyMode
self._alpha = alpha
self._delta = delta
def apply(self, config: dai.RawStereoDepthConfig) -> None:
config.postProcessing.temporalFilter.enable = self._enable
config.postProcessing.temporalFilter.persistencyMode = self._persistencyMode
config.postProcessing.temporalFilter.alpha = self._alpha
config.postProcessing.temporalFilter.delta = self._delta
class SpatialFilter(StereoDepthPostFilter):
"""
Spatial Edge-Preserving Filter will fill invalid depth pixels with valid neighboring depth pixels. It performs a
series of 1D horizontal and vertical passes or iterations, to enhance the smoothness of the reconstructed data.
"""
def __init__(self,
enable: bool = True,
hole_filling_radius: int = 2,
alpha: float = 0.5,
delta: int = 0,
num_iterations: int = 1):
"""
:param enable: Whether to enable or disable the filter.
:param hole_filling_radius: An in-place heuristic symmetric hole-filling mode applied horizontally during
the filter passes. Intended to rectify minor artefacts with minimal performance impact. Search radius for
hole filling.
:param alpha: The Alpha factor in an exponential moving average with Alpha=1 - no filter.
Alpha = 0 - infinite filter. Determines the amount of smoothing.
:param delta: Step-size boundary. Establishes the threshold used to preserve “edges”. If the disparity value
between neighboring pixels exceed the disparity threshold set by this delta parameter, then filtering will be
temporarily disabled. Default value 0 means auto: 3 disparity integer levels. In case of subpixel mode its
3*number of subpixel levels.
:param num_iterations: Number of iterations over the image in both horizontal and vertical direction.
"""
self._enable = enable
self._hole_filling_radius = hole_filling_radius
self._alpha = alpha
self._delta = delta
self._num_iterations = num_iterations
def apply(self, config: dai.RawStereoDepthConfig) -> None:
config.postProcessing.spatialFilter.enable = self._enable
config.postProcessing.spatialFilter.holeFillingRadius = self._hole_filling_radius
config.postProcessing.spatialFilter.alpha = self._alpha
config.postProcessing.spatialFilter.delta = self._delta
config.postProcessing.spatialFilter.numIterations = self._num_iterations
class ThresholdFilter(StereoDepthPostFilter):
"""
Threshold Filter filters out all disparity/depth pixels outside the configured min/max threshold values.
"""
def __init__(self, min_range: int = 400, max_range: int = 15000):
"""
:param min_range: Minimum range in depth units. Depth values under this value are invalidated.
:param max_range: Maximum range in depth units. Depth values over this value are invalidated.
"""
self._min_range = min_range
self._max_range = max_range
def apply(self, config: dai.RawStereoDepthConfig) -> None:
config.postProcessing.thresholdFilter.minRange = self._min_range
config.postProcessing.thresholdFilter.maxRange = self._max_range
class DecimationFilter(StereoDepthPostFilter):
"""
Decimation Filter will sub-samples the depth map, which means it reduces the depth scene complexity and allows
other filters to run faster. Setting decimationFactor to 2 will downscale 1280x800 depth map to 640x400.
"""
def __init__(self,
decimation_factor: int = 1,
decimation_mode: dai.RawStereoDepthConfig.PostProcessing.DecimationFilter.DecimationMode = dai.RawStereoDepthConfig.PostProcessing.DecimationFilter.DecimationMode.PIXEL_SKIPPING
):
"""
:param decimation_factor: Decimation factor. Valid values are 1,2,3,4. Disparity/depth map x/y resolution will
be decimated with this value.
:param decimation_mode: Decimation algorithm type.
"""
self._decimation_factor = decimation_factor
self._mode = decimation_mode
def apply(self, config: dai.RawStereoDepthConfig) -> None:
config.postProcessing.decimationFilter.decimationFactor = self._decimation_factor
config.postProcessing.decimationFilter.decimationMode = self._mode
class DepthSource(Source):
def __init__(self, pipeline: dai.Pipeline,
extended_disparity: bool = False,
subpixel: bool = False,
lr_check: bool = True,
stereo_filters: List[StereoDepthPostFilter] = []
) -> None:
"""
# Closer-in minimum depth, disparity range is doubled (from 95 to 190):
extended_disparity = False
# Better accuracy for longer distance, fractional disparity 32-levels:
subpixel = False
# Better handling for occlusions:
lr_check = True
"""
self._monoLeft = pipeline.create(dai.node.MonoCamera)
self._monoRight = pipeline.create(dai.node.MonoCamera)
self._depth = pipeline.create(dai.node.StereoDepth)
self._xout_disparity = pipeline.create(dai.node.XLinkOut)
self._xout_disparity.setStreamName("disparity")
# Properties
self._monoLeft.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
self._monoLeft.setCamera("left")
self._monoLeft.out.link(self._depth.left)
self._monoRight.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
self._monoRight.setCamera("right")
self._monoRight.out.link(self._depth.right)
# Create a node that will produce the depth map
# (using disparity output as it's easier to visualize depth this way)
self._depth.setDefaultProfilePreset(dai.node.StereoDepth.PresetMode.HIGH_DENSITY)
# Options: MEDIAN_OFF, KERNEL_3x3, KERNEL_5x5, KERNEL_7x7 (default)
self._depth.initialConfig.setMedianFilter(dai.MedianFilter.KERNEL_7x7)
self._depth.setLeftRightCheck(lr_check)
self._depth.setExtendedDisparity(extended_disparity)
self._depth.setSubpixel(subpixel)
self._depth.disparity.link(self._xout_disparity.input)
if len(stereo_filters) > 0:
# Configure post-processing filters
config = self._depth.initialConfig.get()
for filter in stereo_filters:
filter.apply(config)
self._depth.initialConfig.set(config)
def get_stream_name(self) -> str:
return self._xout_disparity.getStreamName()
def link(self, input_node: dai.Node.Input) -> None:
self._depth.disparity.link(input_node)
@dataclass
class MqttConfig:
"""MQTT configuration"""
host: str
topic: str
port: int = 1883
qos: int = 0
class MqttSource(Source):
"""Image source based onto mqtt stream"""
def __init__(self, device: Device, pipeline: dai.Pipeline, mqtt_config: MqttConfig):
self._mqtt_config = mqtt_config
self._client = mqtt.Client()
self._client.user_data_set(mqtt_config)
self._client.on_connect = self._on_connect
self._client.on_message = self._on_message
self._img_in = pipeline.createXLinkIn()
self._img_in.setStreamName("img_input")
self._img_out = pipeline.createXLinkOut()
self._img_out.setStreamName("img_output")
self._img_in.out.link(self._img_out.input)
self._img_in_queue = device.getInputQueue(self._img_in.getStreamName())
def run(self) -> None:
""" Connect and start mqtt loop """
self._client.connect(host=self._mqtt_config.host, port=self._mqtt_config.port)
self._client.loop_start()
def stop(self) -> None:
"""Stop and disconnect mqtt loop"""
self._client.loop_stop()
self._client.disconnect()
@staticmethod
# pylint: disable=unused-argument
def _on_connect(client: mqtt.Client, userdata: MqttConfig, flags: Any,
result_connection: Any) -> None:
# if we lose the connection and reconnect then subscriptions will be renewed.
client.subscribe(topic=userdata.topic, qos=userdata.qos)
# pylint: disable=unused-argument
def _on_message(self, _: mqtt.Client, user_data: MqttConfig, msg: mqtt.MQTTMessage) -> None:
frame_msg = evt.FrameMessage()
frame_msg.ParseFromString(msg.payload)
frame = np.asarray(frame_msg.frame, dtype="uint8")
frame = cv2.imdecode(frame, cv2.IMREAD_COLOR)
nn_data = dai.NNData()
nn_data.setLayer("data", _to_planar(frame, (300, 300)))
self._img_in_queue.send(nn_data)
def get_stream_name(self) -> str:
return self._img_out.getStreamName()
def link(self, input_node: dai.Node.Input) -> None:
self._img_in.out.link(input_node)
def _to_planar(arr: npt.NDArray[np.uint8], shape: tuple[int, int]) -> list[int]:
return [val for channel in cv2.resize(arr, shape).transpose(2, 0, 1) for y_col in channel for val in y_col]
class PipelineController:
"""
Pipeline controller that drive camera device
"""
def __init__(self, frame_processor: FrameProcessor,
object_processor: ObjectProcessor, disparity_processor: DisparityProcessor,
camera: Source, depth_source: Source, object_node: ObjectDetectionNN,
pipeline: dai.Pipeline):
self._frame_processor = frame_processor
self._object_processor = object_processor
self._disparity_processor = disparity_processor
self._camera = camera
self._depth_source = depth_source
self._object_node = object_node
self._stop = False
self._pipeline = pipeline
self._configure_pipeline()
self._focal_length_in_pixels: float | None = None
def _configure_pipeline(self) -> None:
logger.info("configure pipeline")
self._pipeline.setOpenVINOVersion(version=dai.OpenVINO.VERSION_2021_4)
# Link preview to manip and manip to nn
self._camera.link(self._object_node.get_input())
logger.info("pipeline configured")
def run(self) -> None:
"""
Start event loop
:return:
"""
# Connect to device and start pipeline
with Device(pipeline=self._pipeline) as dev:
logger.info('MxId: %s', dev.getDeviceInfo().getMxId())
logger.info('USB speed: %s', dev.getUsbSpeed())
logger.info('Connected cameras: %s', str(dev.getConnectedCameras()))
logger.info("output queues found: %s", str(''.join(dev.getOutputQueueNames()))) # type: ignore
calib_data = dev.readCalibration()
intrinsics = calib_data.getCameraIntrinsics(dai.CameraBoardSocket.CAM_C)
self._focal_length_in_pixels = intrinsics[0][0]
logger.info('Right mono camera focal length in pixels: %s', self._focal_length_in_pixels)
dev.startPipeline()
# Queues
queue_size = 4
q_rgb = dev.getOutputQueue(name=self._camera.get_stream_name(), maxSize=queue_size, # type: ignore
blocking=False)
q_nn = dev.getOutputQueue(name=self._object_node.get_stream_name(), maxSize=queue_size, # type: ignore
blocking=False)
q_disparity = dev.getOutputQueue(name=self._depth_source.get_stream_name(), maxSize=queue_size, # type: ignore
blocking=False)
start_time = time.time()
counter = 0
fps = 0
display_time = time.time()
self._stop = False
while True:
if self._stop:
logger.info("stop loop event")
return
try:
self._loop_on_camera_events(q_nn, q_rgb, q_disparity)
# pylint: disable=broad-except # bad frame or event must not stop loop
except Exception as ex:
logger.exception("unexpected error: %s", str(ex))
counter += 1
if (time.time() - start_time) > 1:
fps = counter / (time.time() - start_time)
counter = 0
start_time = time.time()
if (time.time() - display_time) >= 10:
display_time = time.time()
logger.info("fps: %s", fps)
def _loop_on_camera_events(self, q_nn: dai.DataOutputQueue, q_rgb: dai.DataOutputQueue, q_disparity: dai.DataOutputQueue) -> None:
logger.debug("wait for new frame")
# Wait for frame
in_rgb: dai.ImgFrame = q_rgb.get() # type: ignore # blocking call, will wait until a new data has arrived
try:
logger.debug("process frame")
frame_ref = self._frame_processor.process(in_rgb)
except FrameProcessError as ex:
logger.error("unable to process frame: %s", str(ex))
return
logger.debug("frame processed")
logger.debug("wait for nn response")
# Read NN result
in_nn: dai.NNData = q_nn.get() # type: ignore
logger.debug("process objects")
self._object_processor.process(in_nn, frame_ref)
logger.debug("objects processed")
logger.debug("process disparity")
in_disparity: dai.ImgFrame = q_disparity.get() # type: ignore
self._disparity_processor.process(in_disparity, frame_ref=frame_ref,
focal_length_in_pixels=self._focal_length_in_pixels)
logger.debug("disparity processed")
def stop(self) -> None:
"""
Stop event loop, if loop is not running, do nothing
:return:
"""
self._stop = True
def _bbox_to_object(bbox: npt.NDArray[np.float64], score: float) -> evt.Object:
obj = evt.Object()
obj.type = evt.TypeObject.ANY
obj.top = bbox[0].astype(float)
obj.right = bbox[3].astype(float)
obj.bottom = bbox[2].astype(float)
obj.left = bbox[1].astype(float)
obj.confidence = score
return obj