From c22c92cae374ce342d555d55e1811ec4e95a41f9 Mon Sep 17 00:00:00 2001 From: TPP-G6 Date: Sun, 1 Mar 2026 12:37:59 +0800 Subject: [PATCH] add ffmpeg backend for linux compatibility --- src/orisys/sensor.py | 159 ++++++++++++++++++++++++++++++++++++++----- test_cam.py | 75 ++++++++++++++++++++ test_ffmpeg.py | 46 +++++++++++++ 3 files changed, 264 insertions(+), 16 deletions(-) create mode 100644 test_cam.py create mode 100644 test_ffmpeg.py diff --git a/src/orisys/sensor.py b/src/orisys/sensor.py index a0561a6..b705727 100644 --- a/src/orisys/sensor.py +++ b/src/orisys/sensor.py @@ -7,9 +7,11 @@ and Helmholtz-Hodge decomposition. """ import os +import platform import numpy as np import cv2 import time +import threading from enum import Enum, auto from typing import Tuple # from threading import Thread, Event @@ -21,6 +23,10 @@ from .HHD import nHHD from . import configs from .get_depth import Contact3D from .backend_availability import (check_cupy_available, check_opencl_available) +try: + import av # Optional: used for Linux camera capture path. +except ImportError: + av = None class Sensor: """ @@ -128,6 +134,18 @@ class Sensor: self.__contact3d = Contact3D(config_contact=self.__config_contact, verbose=self.verbose) self.__hhd = None self.poisson_backend = poisson_backend + self._use_pyav = False + self._av_container = None + self._av_stream = None + self._av_decoder = None + self.cap = None + self._is_live_camera = isinstance(self.vid_src, int) + self._capture_thread = None + self._capture_running = False + self._latest_frame = None + self._latest_frame_seq = 0 + self._last_consumed_seq = -1 + self._latest_frame_lock = threading.Lock() # Setup video capture self._open_video_source(override, box) @@ -157,7 +175,7 @@ class Sensor: # Simple direct read - buffer size of 1 should minimize frame accumulation # Timing measurement helps identify if capture or processing is the bottleneck t_capture_start = time.perf_counter() - ret, frame = self.cap.read() + ret, frame = self._read_frame() t_capture_end = time.perf_counter() self._last_capture_time = t_capture_end - t_capture_start @@ -350,7 +368,20 @@ class Sensor: def disconnect(self): """Release resources and stop async threads.""" - if self.cap.isOpened(): + self._capture_running = False + if self._capture_thread is not None: + self._capture_thread.join(timeout=1.0) + self._capture_thread = None + + if self._use_pyav: + if self._av_container is not None: + self._av_container.close() + self._av_container = None + self._av_stream = None + self._av_decoder = None + if self.verbose: + print("Sensor disconnected.") + elif self.cap is not None and self.cap.isOpened(): self.cap.release() if self.verbose: print("Sensor disconnected.") @@ -420,29 +451,57 @@ class Sensor: def _open_video_source(self, override, box): """Initialize video capture and calibration.""" + use_linux_camera_pyav = ( + platform.system() == "Linux" + and isinstance(self.vid_src, int) + and av is not None + ) - self.cap = cv2.VideoCapture(self.vid_src) - # Set buffer size to 1 to reduce latency and avoid frame accumulation - # Note: Not all backends support this, but it helps when available - try: - self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 1) - except: - pass # Some backends don't support this property - self.cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter.fourcc('M', 'J', 'P', 'G')) - self.cap.set(cv2.CAP_PROP_FPS, self._config_fps) - self.cap.set(cv2.CAP_PROP_FRAME_WIDTH, self._config_resolution_raw[0]) - self.cap.set(cv2.CAP_PROP_FRAME_HEIGHT, self._config_resolution_raw[1]) + if use_linux_camera_pyav: + device = f"/dev/video{self.vid_src}" + try: + self._av_container = av.open( + device, + format="v4l2", + options={ + "input_format": "mjpeg", + "video_size": f"{self._config_resolution_raw[0]}x{self._config_resolution_raw[1]}", + "framerate": str(self._config_fps), + }, + ) + self._av_stream = self._av_container.streams.video[0] + self._av_decoder = self._av_container.decode(self._av_stream) + self._use_pyav = True + if self.verbose: + print(f"Using PyAV Linux camera backend: {device}") + except Exception as exc: + if self.verbose: + print(f"PyAV camera open failed ({exc}), fallback to OpenCV backend.") + self._use_pyav = False + + if not self._use_pyav: + self.cap = cv2.VideoCapture(self.vid_src) + # Set buffer size to 1 to reduce latency and avoid frame accumulation + # Note: Not all backends support this, but it helps when available + try: + self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 1) + except Exception: + pass # Some backends don't support this property + self.cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter.fourcc('M', 'J', 'P', 'G')) + self.cap.set(cv2.CAP_PROP_FPS, self._config_fps) + self.cap.set(cv2.CAP_PROP_FRAME_WIDTH, self._config_resolution_raw[0]) + self.cap.set(cv2.CAP_PROP_FRAME_HEIGHT, self._config_resolution_raw[1]) if self.verbose: print("Connecting to video source...") - if not self.cap.isOpened(): + if not self._is_source_open(): raise Exception("Failed to open video source.") # Wait for first valid frame frame = None while frame is None: - ret, frame = self.cap.read() + ret, frame = self._read_frame_raw() if not ret: time.sleep(0.1) @@ -463,6 +522,10 @@ class Sensor: self.initial_img = cv2.flip(self.initial_img, 1) self.initial_img = self.initial_img.astype(np.uint8) + # For live cameras, start a latest-frame capture thread to avoid latency buildup. + if self._is_live_camera: + self._start_capture_thread() + def _find_stitch_boxes(self, img_src, show=False): """Detect vision tile boxes with camera stabilization.""" @@ -470,7 +533,9 @@ class Sensor: start_time = cv2.getTickCount() while n_boxes < self._config_units: - ret, frame = self.cap.read() + ret, frame = self._read_frame_raw() + if not ret: + continue v_mean = np.sum(cv2.mean(frame)) if abs(v_mean - v_mean0) < 5: @@ -496,6 +561,68 @@ class Sensor: self.stitcher.extract_tiles() return True + def _is_source_open(self): + if self._use_pyav: + return self._av_container is not None + return self.cap is not None and self.cap.isOpened() + + def _read_frame(self): + if self._capture_running: + # Startup race guard: wait briefly until capture thread publishes first frame. + if self._latest_frame is None: + deadline = time.perf_counter() + 1.0 + while time.perf_counter() < deadline: + with self._latest_frame_lock: + if self._latest_frame is not None: + break + time.sleep(0.001) + + # Return latest frame immediately (do not wait for next camera tick). + # This keeps end-to-end latency minimal; caller may receive repeated + # frames if processing runs faster than camera FPS. + with self._latest_frame_lock: + frame = self._latest_frame + seq = self._latest_frame_seq + if frame is not None: + self._last_consumed_seq = seq + return True, frame.copy() + return False, None + return self._read_frame_raw() + + def _read_frame_raw(self): + if self._use_pyav: + if self._av_decoder is None: + return False, None + try: + frame = next(self._av_decoder) + except StopIteration: + return False, None + except Exception: + return False, None + return True, frame.to_ndarray(format="bgr24") + if self.cap is None: + return False, None + return self.cap.read() + + def _start_capture_thread(self): + if self._capture_thread is not None: + return + self._capture_running = True + self._capture_thread = threading.Thread(target=self._capture_loop, daemon=True) + self._capture_thread.start() + if self.verbose: + print("Latest-frame capture thread started.") + + def _capture_loop(self): + while self._capture_running: + ret, frame = self._read_frame_raw() + if not ret: + time.sleep(0.001) + continue + with self._latest_frame_lock: + self._latest_frame = frame + self._latest_frame_seq += 1 + # ========================================================================= # Private - Utilities diff --git a/test_cam.py b/test_cam.py new file mode 100644 index 0000000..6f5e2c7 --- /dev/null +++ b/test_cam.py @@ -0,0 +1,75 @@ +import cv2 +import time + +# Toggle these to test whether exposure is limiting delivered FPS. +USE_MANUAL_EXPOSURE = False +MANUAL_EXPOSURE = -6 # Typical V4L2/OpenCV manual range is negative values. + +cap = cv2.VideoCapture(4, cv2.CAP_V4L2) +cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter.fourcc('M', 'J', 'P', 'G')) +cap.set(cv2.CAP_PROP_FPS, 30) +cap.set(cv2.CAP_PROP_FRAME_WIDTH, 1920) +cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 1080) +cap.set(cv2.CAP_PROP_BUFFERSIZE, 1) + +# For V4L2 in OpenCV: auto-exposure is usually 3, manual is usually 1. +if USE_MANUAL_EXPOSURE: + ok_auto = cap.set(cv2.CAP_PROP_AUTO_EXPOSURE, 1) + ok_exp = cap.set(cv2.CAP_PROP_EXPOSURE, MANUAL_EXPOSURE) +else: + ok_auto = cap.set(cv2.CAP_PROP_AUTO_EXPOSURE, 3) + ok_exp = True + +actual_fourcc = int(cap.get(cv2.CAP_PROP_FOURCC)) +print(f"Backend: {cap.getBackendName()}") +print( + f"FOURCC: {chr(actual_fourcc & 0xFF)}" + f"{chr((actual_fourcc >> 8) & 0xFF)}" + f"{chr((actual_fourcc >> 16) & 0xFF)}" + f"{chr((actual_fourcc >> 24) & 0xFF)}" +) +print(f"Resolution: {int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))}x{int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))}") +print(f"FPS: {cap.get(cv2.CAP_PROP_FPS)}") +print( + f"Exposure mode: {'manual' if USE_MANUAL_EXPOSURE else 'auto'} " + f"(set_auto_ok={ok_auto}, set_exp_ok={ok_exp})" +) +print(f"AutoExposure(readback): {cap.get(cv2.CAP_PROP_AUTO_EXPOSURE)}") +print(f"Exposure(readback): {cap.get(cv2.CAP_PROP_EXPOSURE)}") + +frame_count = 0 +start = time.perf_counter() +while cap.isOpened(): + t0 = time.perf_counter() + ok = cap.grab() + t1 = time.perf_counter() + if not ok: + break + + ret, frame = cap.retrieve() + t2 = time.perf_counter() + if not ret: + continue + + cv2.imshow("webcam", frame) + t3 = time.perf_counter() + key = cv2.waitKey(1) & 0xFF + t4 = time.perf_counter() + + frame_count += 1 + if frame_count % 30 == 0: + elapsed = t4 - start + print( + f"Frame {frame_count}: " + f"grab={1000*(t1-t0):.1f}ms, " + f"retrieve={1000*(t2-t1):.1f}ms, " + f"imshow={1000*(t3-t2):.1f}ms, " + f"waitKey={1000*(t4-t3):.1f}ms, " + f"avg_fps={frame_count/elapsed:.1f}" + ) + + if key == ord('q'): + break + +cap.release() +cv2.destroyAllWindows() \ No newline at end of file diff --git a/test_ffmpeg.py b/test_ffmpeg.py new file mode 100644 index 0000000..8656e46 --- /dev/null +++ b/test_ffmpeg.py @@ -0,0 +1,46 @@ +import av +import cv2 +import time + +DEVICE = "/dev/video4" +TARGET_FPS = 30 +WIDTH = 1280 +HEIGHT = 720 + +# Open v4l2 device with same style options as ffmpeg CLI +container = av.open( + DEVICE, + format="v4l2", + options={ + "input_format": "mjpeg", + "video_size": f"{WIDTH}x{HEIGHT}", + "framerate": str(TARGET_FPS), + }, +) + +stream = container.streams.video[0] +frame_count = 0 +t_start = time.perf_counter() + +for frame in container.decode(stream): + t0 = time.perf_counter() + img = frame.to_ndarray(format="bgr24") # OpenCV-compatible + t1 = time.perf_counter() + + cv2.imshow("pyav_cam", img) + key = cv2.waitKey(1) & 0xFF + t2 = time.perf_counter() + + frame_count += 1 + if frame_count % 30 == 0: + elapsed = t2 - t_start + print( + f"Frame {frame_count}: convert={1000*(t1-t0):.1f}ms, " + f"imshow+waitKey={1000*(t2-t1):.1f}ms, avg_fps={frame_count/elapsed:.1f}" + ) + + if key == ord("q"): + break + +container.close() +cv2.destroyAllWindows() \ No newline at end of file