Hi Team,
We are building a product for Indian Government. We have purchased Zed2i Camera. We are trying to measure distance of pole from moving vehicle. Zed2i is placed on a moving vehicle. We are using object detection model to first detect pole and we are doing great at it. We are getting perfect bounding boxes for poles. Here are the setting we are using for Depth:
zed = sl.Camera()
zed.set_camera_settings(sl.VIDEO_SETTINGS.EXPOSURE, -1)
zed.set_camera_settings(sl.VIDEO_SETTINGS.GAIN, -1)
init_params = sl.InitParameters(
depth_mode=sl.DEPTH_MODE.ULTRA,
coordinate_units=sl.UNIT.INCH,
camera_resolution=sl.RESOLUTION.HD1080,
camera_fps=30,
depth_maximum_distance=780,
depth_minimum_distance=8,
depth_stabilization = 1
)
init_params.camera_resolution = sl.RESOLUTION.HD1080
init_params.camera_fps = 30
We are getting distance error in the range 3 to 4 inches always. Weather is sunny and objects is perfectly visible.
Here is how we are getting depth values from bounding box:
def get_depth_values_in_roi(self, x1, y1, x2, y2):
with self.device:
depth_array = cp.array(self.depth)
roi_depth_values = depth_array[y1:y2, x1:x2]
valid_mask = (roi_depth_values > 0) & (roi_depth_values < self.MAX_DETECTION_DISTANCE)
valid_depth_values = roi_depth_values[valid_mask]
return valid_depth_values
We are taking median of those valid depth values to get final distance number:
def process_depth_values(self, depth_values):
if len(depth_values) == 0:
return float('inf') # Return a large value if no valid depths are found
return float(cp.median(depth_values))
I am attaching image containing target object.
Please suggest:
- Best settings to measure depth correctly. Contraint here is I need to use 30FPS atleast. Might even switch to 60 FPS at 720 due to vehicle speed.
- Also, methods to get only target object depths only. Can not use segmentation models here due to system limitation.
Thank you for reaching out.