Skip to content

v5.5.0

Choose a tag to compare

@Myzhar Myzhar released this 18 Sep 14:42
· 23 commits to master since this release

v5.5.0

  • Fixed the large IMU publishing delay caused by the IMU being read in step with grab() (stereo and mono). With general.grab_frame_rate: 15 and general.grab_compute_capping_fps: 5 the samples arrived in 5 Hz bursts several hundred milliseconds late; the mono node, which has no compute capping, was delayed by the grab rate alone. Both nodes now read the sensors stream independently of grab(), polling well above the hardware ODR so that no sample is lost. Measured on a ZED X on an AGX Orin: median publish delay 411 ms -> 10.5 ms, output rate 92.6 -> 100.0 Hz, and the delay no longer depends on general.grab_compute_capping_fps. SVO and simulation are unchanged.

  • Fixed sensors.sensors_pub_rate not being honored on cameras whose IMU does not run at the rate the ZED SDK advertises (stereo and mono). A ZED X One GS reports 400 Hz but delivers 200 Hz, so sensors_pub_rate: 100 produced a steady 50 Hz. Both nodes now decimate against the measured sample interval instead of the advertised one. Cameras that report correctly now produce exactly the requested rate rather than the nearest sub-multiple of the real one (a ZED 2i publishes 100.0 Hz instead of 101.1 Hz).

  • Reduced the CPU usage of both nodes by counting topic subscribers through the publisher instead of a ROS graph query, and by refreshing the sensors thread counts every 250 ms instead of on every poll. On an AGX Orin this cuts the sensors thread overhead from ~5% to ~2-3% of a CPU core, whether or not the IMU topics are subscribed. Two consequences: a new subscriber to a sensors topic can wait up to 250 ms for its first message, and a subscriber whose QoS is incompatible with the publisher no longer triggers publishing (it could never receive the data anyway).

  • Added the general.bus_type parameter to the ZedCameraOne node (AUTO default, USB, GMSL, and with ZED SDK >= 5.5 MIPI and HOLOSCAN), used when the camera is selected with general.camera_id. The bus was hard-coded to GMSL, so a ZED X One Core on a MIPI capture card, or reached over Camera-over-Ethernet through a Holoscan sensor bridge, could not be opened by ID at all.

  • Added the pause_svo_rec service (stereo and mono, std_srvs/srv/SetBool) to pause and resume an SVO recording without closing the file. It returns success: false when no recording is active, and the diagnostic reports SVO Recording: PAUSED while paused.

  • Added the general.max_system_clock_step_ms parameter (stereo and mono) to set how fast the SDK follows a backward host-clock adjustment (NTP/PTP step, manual date change): 4.0 is the ZED SDK default and suits slew-only setups, a negative value makes step corrections land instantly as PTP deployments expect, and 0.0 freezes the offset captured on first use. Applied only when general.sdk_timestamp_clock is SYSTEM_CLOCK. Requires ZED SDK >= 5.3.

  • Added the general.sdk_timestamp_clock parameter (stereo and mono) to select the clock behind every SDK timestamp: SYSTEM_CLOCK (default), MONOTONIC_CLOCK, or MONOTONIC_RAW_CLOCK (ZED SDK >= 5.4), which is immune to NTP/PTP frequency slewing as well as to clock steps. It replaces the boolean general.sdk_use_monotonic_clock, which is still read and still works, but now logs a deprecation warning.

  • Added the svo.encryption_key parameter (stereo and mono) to encrypt the SVO files created by start_svo_rec (AES-256-CTR); playback was already supported through svo.decryption_key. It accepts a path to a 32-byte key file, a 64-character hex key or a passphrase, and is never written to the log. Requires ZED SDK >= 5.3 and OpenSSL at runtime.

  • Added the positional tracking pose confidence, a [0,100] quality metric filled by POSITIONAL_TRACKING_MODE::GEN_3 where 0 means tracking is lost. It is reported as Pose confidence in the diagnostic and published as the new pose_confidence field of PosTrackStatus on ~/pose/status. Requires ZED SDK >= 5.4 and zed_msgs v5.5.

  • Fixed the distortion_model and the distortion coefficients published in the camera_info topics. They were chosen from a hard-coded per-model table that the lens does not follow: a fisheye ZED X One was published as rational_polynomial instead of equidistant, and the ZED X HDR models were missing from the table entirely, so their camera_info carried no distortion model at all. Both nodes now read the model from the ZED SDK (>= 5.4) and fall back to the old heuristic on older versions. Rectified parameters are reported as plumb_bob with zero coefficients.

  • Aligned the zed_debug configuration files with the zed_wrapper ones: five parameters were missing and custom_object_detection.yaml lagged four per-class tracking parameters behind. Both directories now declare the same parameters; only the debug-specific values differ.

  • Fixed general.sdk_use_monotonic_clock sitting in the debug section of zed_wrapper/config/common_mono.yaml and zed_debug/config/common_stereo.yaml, where it was read under the wrong name and silently ignored.

  • Added the missing advanced.change_thread_priority to common_stereo.yaml and to both common_mono.yaml files. It gates every other advanced.* parameter and defaults to false, so the thread scheduling policy and priorities listed next to it were never applied unless it was added by hand.

  • Fixed the sports ball class in custom_object_detection.yaml, whose label was corrupted to pesports ballrson by a bad search and replace. The other 79 labels and every model_class_id were verified against the COCO 2017 order.

  • Improved the per-class defaults of custom_object_detection.yaml: is_grounded is now set only for objects that really lie on the floor (it was set on tableware, food and laptops, whose 3D position was then projected onto the floor) and is_static only for objects fixed in the world, and the tracking timeout, 3D box bounds, velocity settings and confirmation and prediction times are now set per object type instead of being identical for all 80 classes.

  • Fixed object_detection.class_XXX.min_box_height_meters being documented for every class but never read by the node, so the minimum 3D height was never applied.

  • Added CUSTOM_RFDETRLIKE_BOX_OBJECTS and CUSTOM_BOX_OBJECTS_AUTODETECT as valid object_detection.detection_model values (ZED SDK >= 5.5). They were accepted by the parameter parser but treated as built-in models, so the ONNX file was never passed to the ZED SDK and enabling object detection failed.

  • Fixed the ONNX input shape sent to the ZED SDK for custom object detection models: object_detection.custom_onnx_input_size produced a size x 0 resolution instead of a square one, so the shape never matched the one a model had been optimized with and the engine had to be rebuilt.

  • Added the camera health status to the diagnostic output: image quality, lighting, depth reliability, motion sensors reliability and, with ZED SDK >= 5.5 and zed_msgs v5.5, duplicated images. The diagnostic goes to WARN and names the issues while any of them is raised, and reports DISABLED when general.enable_image_validity_check is off.

  • Added the duplicated_image field to the HealthStatusStamped message published on ~/status/health, raised by the ZED SDK when a frame repeats the content of an earlier one under a new timestamp, i.e. a repeating or stalled stream. Requires ZED SDK >= 5.5 and zed_msgs v5.5; with an older zed_msgs the node runs as before, without the field.

  • Added the general.timestamp_reference parameter (IMAGE default, IMAGE_CENTER_OF_EXPOSURE) to both nodes. IMAGE_CENTER_OF_EXPOSURE stamps the published data with the middle of the frame exposure instead of the start of the sensor readout, which is what aligns frames with sensors timestamped at the instant they measure (LiDAR, robot joints). It needs an input carrying a per-frame exposure (ZED X, ZED X Mini, ZED X One GS, ZED X One 4K): on USB and HDR models the node warns once and falls back to IMAGE. Requires ZED SDK >= 5.5.

  • Changed the default body_tracking.body_format from BODY_38 to BODY_34, and body_tracking.enable_body_fitting from false to true. Only BODY_18 and BODY_34 have a GEN_2 network, so the previous default silently fell back to the older GEN_1. Skeletons published on ~/body_trk/skeletons now carry 34 keypoints instead of 38: set body_format back to BODY_38 to keep the previous topology, at the cost of running GEN_1.

  • Added the body_tracking.model_gen parameter (GEN_1, GEN_2; default GEN_2) to select the body tracking network. GEN_2 is more robust in crowded scenes and to unusual body orientations, detects people further from the camera, and is up to 18% faster and 14% more accurate than GEN_1. body_tracking.allow_reduced_precision_inference has no effect on GEN_2, which always runs in FP16. Requires ZED SDK >= 5.5.

  • Added the pos_tracking.compute_preference parameter (AUTO default, PREFER_CPU, PREFER_GPU) to control how much GPU the positional tracking may use: PREFER_GPU makes tracking faster and shortens grab(), PREFER_CPU keeps the GPU free for your own workloads. GEN_1 computes depth and so uses the GPU whatever the setting. Requires ZED SDK >= 5.5.

  • Added the depth.allow_depth_cuda_graph parameter (default false): the ZED SDK records the depth computation once and replays it at every grab, lowering the CPU cost of launching it, which matters most on Jetson. The depth output is unchanged. It applies to the NEURAL depth mode family only, and the SDK reverts to the regular computation if the recording fails. Requires ZED SDK >= 5.5.

  • Added the depth.depth_precision parameter to select the precision of the neural depth inference: FP16 (default, unchanged behavior) or INT8, which is faster and uses less GPU memory for a small accuracy cost. INT8 requires depth.depth_mode: 'NEURAL'; the ZED SDK falls back to FP16 when the depth mode or the GPU does not support it. Requires ZED SDK >= 5.5.

  • Fixed a zed_msgs version check that blocked scene_illuminance information in health status messages.

  • Fixed scene_illuminance reporting a value frozen at start-up in both the health status message and the Scene Illuminance diagnostic. It is now refreshed while grabbing and tracks the actual scene light level.

  • Improved simulation mode to match the features of the ZED extension for NVIDIA Isaac Sim:

    • The grab resolution and framerate are no longer forced to HD1080 @ 60 Hz, but read from the simulated camera, so a simulation running at the extension default of 30 FPS is reported and published correctly. This also removes the bogus "general.grab_frame_rate value is not valid" warning printed at every start-up, and the capture diagnostics reporting half the expected rate.
    • general.pub_frame_rate: 0 ("no limit") now follows the real grab rate even when it is only known after the input is open.
    • Fixed the unstable IMU publishing rate in simulation: the stereo node now drains the IMU FIFO as the mono node already did, instead of polling only the latest sample. (Live cameras no longer take this path: see the IMU delay entry above.)
    • Fixed the simulation connection timeout never expiring when use_sim_time:=true and no /clock publisher was running yet. The timeout error now reports the configured simulation.sim_address / simulation.sim_port to check.
    • A simulated Virtual Stereo Camera (camera_model:=virtual) no longer requires serial_numbers or camera_ids: the simulator streams the already paired stereo images.
    • The "system overloaded" diagnostic now points to the simulator FPS / Resolution fields instead of node parameters that have no effect in simulation.
    • Silenced the spurious SENSORS_NOT_AVAILABLE warnings produced by a simulated camera without an IMU.
  • Added simulation mode support for the ZED X One monocular cameras: launch ZedCameraOne with sim_mode:=true (plus the optional sim_address, sim_port and use_sim_time options) to publish images and IMU data from the simulated camera. With use_sim_time the node waits for a valid /clock and stamps every topic with the simulation time.