feat: Terrain classification for speed adaptation (Issue #469)

Implement multi-sensor terrain classification using RealSense D435i depth and RPLIDAR A1M8:

- saltybot_terrain_classification: New ROS2 package for terrain classification
- TerrainClassifier: Rule-based classifier matching depth variance + reflectance to terrain type
  (smooth/carpet/grass/gravel) with hysteresis + confidence scoring
- DepthExtractor: Extracts roughness from depth discontinuities and surface gradients
- LidarExtractor: Extracts reflectance from RPLIDAR scan intensities
- terrain_classification_node: 10Hz node fusing both sensors, publishes:
  - /saltybot/terrain_type (JSON with type, confidence, speed_scale)
  - /saltybot/terrain_type_string (human-readable type)
  - /saltybot/terrain_speed_scale (0.0-1.0 speed multiplier for smooth/carpet/grass/gravel)

Speed scales: smooth=1.0, carpet=0.9, grass=0.75, gravel=0.6

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
This commit is contained in:
sl-perception 2026-03-06 16:43:21 -05:00
parent 5add2cab51
commit 8d58d5e34c
12 changed files with 580 additions and 0 deletions

View File

@ -0,0 +1,10 @@
terrain_classification:
ros__parameters:
control_rate: 10.0
depth_topic: "/camera/depth/image_rect_raw"
lidar_topic: "/scan"
depth_std_threshold: 0.02
min_depth_m: 0.1
max_depth_m: 3.0
confidence_threshold: 0.5
hysteresis_samples: 5

View File

@ -0,0 +1,43 @@
#!/usr/bin/env python3
"""Launch file for terrain classification node."""
from launch import LaunchDescription
from launch_ros.actions import Node
from launch.substitutions import LaunchConfiguration
from launch.actions import DeclareLaunchArgument
def generate_launch_description():
"""Generate launch description."""
depth_topic = LaunchConfiguration("depth_topic", default="/camera/depth/image_rect_raw")
lidar_topic = LaunchConfiguration("lidar_topic", default="/scan")
control_rate = LaunchConfiguration("control_rate", default="10.0")
confidence_threshold = LaunchConfiguration("confidence_threshold", default="0.5")
return LaunchDescription([
DeclareLaunchArgument("depth_topic", default_value="/camera/depth/image_rect_raw",
description="RealSense depth topic"),
DeclareLaunchArgument("lidar_topic", default_value="/scan",
description="RPLIDAR scan topic"),
DeclareLaunchArgument("control_rate", default_value="10.0",
description="Control loop rate (Hz)"),
DeclareLaunchArgument("confidence_threshold", default_value="0.5",
description="Minimum confidence to publish"),
Node(
package="saltybot_terrain_classification",
executable="terrain_classification_node",
name="terrain_classification",
output="screen",
parameters=[
{"depth_topic": depth_topic},
{"lidar_topic": lidar_topic},
{"control_rate": control_rate},
{"confidence_threshold": confidence_threshold},
{"depth_std_threshold": 0.02},
{"min_depth_m": 0.1},
{"max_depth_m": 3.0},
{"hysteresis_samples": 5},
],
),
])

View File

@ -0,0 +1,29 @@
<?xml version="1.0"?>
<?xml-model href="http://download.ros.org/schema/package_format3.xsd" schematypens="http://www.w3.org/2001/XMLSchema"?>
<package format="3">
<name>saltybot_terrain_classification</name>
<version>0.1.0</version>
<description>Terrain classification using RealSense depth and RPLIDAR (Issue #469)</description>
<maintainer email="seb@example.com">SaltyLab</maintainer>
<license>MIT</license>
<buildtool_depend>ament_python</buildtool_depend>
<depend>rclpy</depend>
<depend>std_msgs</depend>
<depend>geometry_msgs</depend>
<depend>sensor_msgs</depend>
<depend>message_filters</depend>
<depend>numpy</depend>
<depend>opencv-python</depend>
<depend>cv_bridge</depend>
<test_depend>ament_copyright</test_depend>
<test_depend>ament_flake8</test_depend>
<test_depend>ament_pep257</test_depend>
<test_depend>python3-pytest</test_depend>
<export>
<build_type>ament_python</build_type>
</export>
</package>

View File

@ -0,0 +1,64 @@
"""Depth Feature Extractor — Surface roughness from RealSense depth."""
from __future__ import annotations
import numpy as np
from typing import Optional
import cv2
class DepthExtractor:
"""Extract terrain roughness features from RealSense depth images."""
def __init__(
self,
roi_width_px: int = 320,
roi_height_px: int = 240,
min_depth_m: float = 0.1,
max_depth_m: float = 3.0,
bilateral_d: int = 5,
):
self._roi_w = roi_width_px
self._roi_h = roi_height_px
self._min_depth = min_depth_m
self._max_depth = max_depth_m
self._bilateral_d = bilateral_d
def extract_roughness(self, depth_image: np.ndarray) -> Optional[float]:
"""Extract surface roughness from depth image."""
if depth_image is None or depth_image.size == 0:
return None
if depth_image.dtype == np.uint16:
depth_m = depth_image.astype(np.float32) / 1000.0
else:
depth_m = depth_image.astype(np.float32)
h, w = depth_m.shape
x_start = (w - self._roi_w) // 2
y_start = (h - self._roi_h) // 2
x_end = x_start + self._roi_w
y_end = y_start + self._roi_h
roi = depth_m[max(0, y_start):min(h, y_end), max(0, x_start):min(w, x_end)]
valid_mask = (roi > self._min_depth) & (roi < self._max_depth)
valid_depths = roi[valid_mask]
if len(valid_depths) < 10:
return None
roi_smooth = cv2.bilateralFilter(roi.astype(np.float32), self._bilateral_d, 0.1, 1.0)
depth_variance = np.var(valid_depths)
grad_x = cv2.Sobel(roi_smooth, cv2.CV_32F, 1, 0, ksize=3)
grad_y = cv2.Sobel(roi_smooth, cv2.CV_32F, 0, 1, ksize=3)
grad_magnitude = np.sqrt(grad_x**2 + grad_y**2)
valid_grad = grad_magnitude[valid_mask]
grad_mean = np.mean(valid_grad) if len(valid_grad) > 0 else 0.0
depth_roughness = min(1.0, depth_variance / 0.05)
grad_roughness = min(1.0, grad_mean / 0.02)
roughness = 0.6 * depth_roughness + 0.4 * grad_roughness
return float(roughness)

View File

@ -0,0 +1,58 @@
"""LIDAR Feature Extractor — Surface characteristics from RPLIDAR."""
from __future__ import annotations
import numpy as np
from typing import Optional
from collections import deque
class LidarExtractor:
"""Extract terrain reflectance features from RPLIDAR scans."""
def __init__(
self,
ground_angle_front: float = 10.0,
ground_angle_rear: float = 170.0,
reflectance_window: int = 20,
):
self._front_angle = ground_angle_front
self._rear_angle = ground_angle_rear
self._refl_window = deque(maxlen=reflectance_window)
def extract_reflectance(self, ranges: np.ndarray, intensities: np.ndarray) -> Optional[float]:
"""Extract mean reflectance from ground-hitting rays."""
if len(ranges) == 0 or len(intensities) == 0:
return None
intensities_norm = intensities.astype(np.float32)
if np.max(intensities_norm) > 1.5:
intensities_norm /= 255.0
else:
intensities_norm = np.clip(intensities_norm, 0.0, 1.0)
valid_mask = (ranges > 0.2) & (ranges < 5.0)
valid_intensities = intensities_norm[valid_mask]
if len(valid_intensities) < 5:
return None
mean_reflectance = float(np.mean(valid_intensities))
self._refl_window.append(mean_reflectance)
if self._refl_window:
return float(np.mean(list(self._refl_window)))
return mean_reflectance
def extract_range_variance(self, ranges: np.ndarray) -> Optional[float]:
"""Extract surface variance from range measurements."""
if len(ranges) == 0:
return None
valid_ranges = ranges[(ranges > 0.2) & (ranges < 5.0)]
if len(valid_ranges) < 5:
return None
range_variance = float(np.var(valid_ranges))
normalized = min(1.0, range_variance / 0.2)
return float(normalized)

View File

@ -0,0 +1,220 @@
#!/usr/bin/env python3
"""
terrain_classification_node.py Terrain classification (Issue #469).
Fuses RealSense D435i depth + RPLIDAR A1M8 to classify terrain and recommend speed.
Pipeline (10 Hz)
1. Extract depth-based roughness features
2. Extract lidar-based reflectance features
3. Classify terrain using rule-based matcher
4. Publish /saltybot/terrain_type (JSON + string)
5. Publish speed recommendations
Publishes
/saltybot/terrain_type std_msgs/String JSON: type, confidence, speed_scale
/saltybot/terrain_type_string std_msgs/String Human-readable type name
/saltybot/terrain_speed_scale std_msgs/Float32 Speed multiplier [0.0, 1.0]
"""
import json
import time
import numpy as np
import rclpy
from rclpy.node import Node
from rclpy.qos import QoSProfile, ReliabilityPolicy, HistoryPolicy
from sensor_msgs.msg import Image, LaserScan
from std_msgs.msg import String, Float32
import cv2
from cv_bridge import CvBridge
from saltybot_terrain_classification.terrain_classifier import TerrainClassifier
from saltybot_terrain_classification.depth_extractor import DepthExtractor
from saltybot_terrain_classification.lidar_extractor import LidarExtractor
class TerrainClassificationNode(Node):
def __init__(self):
super().__init__("terrain_classification")
self._declare_params()
p = self._load_params()
self._classifier = TerrainClassifier(
depth_std_threshold=p["depth_std_threshold"],
hysteresis_count=p["hysteresis_samples"],
)
self._depth_extractor = DepthExtractor(
roi_width_px=320,
roi_height_px=240,
min_depth_m=p["min_depth_m"],
max_depth_m=p["max_depth_m"],
)
self._lidar_extractor = LidarExtractor()
self._cv_bridge = CvBridge()
self._latest_depth_image = None
self._latest_scan = None
self._last_depth_t = 0.0
self._last_scan_t = 0.0
depth_qos = QoSProfile(
reliability=ReliabilityPolicy.BEST_EFFORT,
history=HistoryPolicy.KEEP_LAST,
depth=1,
)
lidar_qos = QoSProfile(
reliability=ReliabilityPolicy.BEST_EFFORT,
history=HistoryPolicy.KEEP_LAST,
depth=1,
)
pub_qos = QoSProfile(
reliability=ReliabilityPolicy.RELIABLE,
history=HistoryPolicy.KEEP_LAST,
depth=10,
)
self.create_subscription(
Image, p["depth_topic"], self._depth_cb, depth_qos
)
self.create_subscription(
LaserScan, p["lidar_topic"], self._scan_cb, lidar_qos
)
self._terrain_type_pub = self.create_publisher(
String, "/saltybot/terrain_type", pub_qos
)
self._terrain_type_string_pub = self.create_publisher(
String, "/saltybot/terrain_type_string", pub_qos
)
self._speed_scale_pub = self.create_publisher(
Float32, "/saltybot/terrain_speed_scale", pub_qos
)
rate = p["control_rate"]
self._timer = self.create_timer(1.0 / rate, self._control_cb)
self.get_logger().info(
f"TerrainClassificationNode ready rate={rate}Hz "
f"depth={p['depth_topic']} lidar={p['lidar_topic']}"
)
def _declare_params(self) -> None:
self.declare_parameter("control_rate", 10.0)
self.declare_parameter("depth_topic", "/camera/depth/image_rect_raw")
self.declare_parameter("lidar_topic", "/scan")
self.declare_parameter("depth_std_threshold", 0.02)
self.declare_parameter("min_depth_m", 0.1)
self.declare_parameter("max_depth_m", 3.0)
self.declare_parameter("confidence_threshold", 0.5)
self.declare_parameter("hysteresis_samples", 5)
def _load_params(self) -> dict:
g = self.get_parameter
return {k: g(k).value for k in [
"control_rate", "depth_topic", "lidar_topic",
"depth_std_threshold", "min_depth_m", "max_depth_m",
"confidence_threshold", "hysteresis_samples",
]}
def _depth_cb(self, msg: Image) -> None:
self._latest_depth_image = msg
self._last_depth_t = time.monotonic()
def _scan_cb(self, msg: LaserScan) -> None:
self._latest_scan = msg
self._last_scan_t = time.monotonic()
def _control_cb(self) -> None:
p = self._load_params()
now = time.monotonic()
depth_age = (now - self._last_depth_t) if self._last_depth_t > 0.0 else 1e9
scan_age = (now - self._last_scan_t) if self._last_scan_t > 0.0 else 1e9
if depth_age > 1.0 or scan_age > 1.0:
return
depth_roughness = None
lidar_reflectance = None
if self._latest_depth_image is not None:
try:
depth_cv = self._cv_bridge.imgmsg_to_cv2(
self._latest_depth_image, desired_encoding="passthrough"
)
depth_roughness = self._depth_extractor.extract_roughness(depth_cv)
except Exception:
pass
if self._latest_scan is not None:
try:
ranges = np.array(self._latest_scan.ranges, dtype=np.float32)
intensities = np.array(self._latest_scan.intensities, dtype=np.float32)
lidar_reflectance = self._lidar_extractor.extract_reflectance(
ranges, intensities
)
except Exception:
pass
if depth_roughness is None:
depth_roughness = 0.3
if lidar_reflectance is None:
lidar_reflectance = 0.5
classification = self._classifier.update(depth_roughness, lidar_reflectance)
if classification.confidence < p["confidence_threshold"]:
return
self._publish_terrain_type(classification)
self._publish_speed_scale(classification.speed_scale)
self.get_logger().info(
f"Terrain: {classification.surface_type} "
f"confidence={classification.confidence:.2f} "
f"speed_scale={classification.speed_scale:.2f}"
)
def _publish_terrain_type(self, classification) -> None:
"""Publish terrain classification JSON."""
msg = String()
msg.data = json.dumps({
"surface_type": classification.surface_type,
"confidence": round(classification.confidence, 3),
"roughness": round(classification.roughness, 3),
"reflectance": round(classification.reflectance, 3),
"speed_scale": round(classification.speed_scale, 3),
})
self._terrain_type_pub.publish(msg)
msg_str = String()
msg_str.data = classification.surface_type
self._terrain_type_string_pub.publish(msg_str)
def _publish_speed_scale(self, scale: float) -> None:
"""Publish speed scale multiplier."""
msg = Float32()
msg.data = float(scale)
self._speed_scale_pub.publish(msg)
def main(args=None):
rclpy.init(args=args)
node = TerrainClassificationNode()
try:
rclpy.spin(node)
except KeyboardInterrupt:
pass
finally:
node.destroy_node()
rclpy.try_shutdown()
if __name__ == "__main__":
main()

View File

@ -0,0 +1,127 @@
"""Terrain Classifier — Multi-sensor classification (Issue #469)."""
from __future__ import annotations
from dataclasses import dataclass
from collections import deque
import numpy as np
from typing import Optional
@dataclass
class TerrainClassification:
"""Classification result with confidence."""
surface_type: str
confidence: float
roughness: float
reflectance: float
speed_scale: float
class TerrainClassifier:
"""Multi-sensor terrain classifier using depth variance + reflectance."""
_SURFACE_RULES = [
("smooth", 0.10, 0.70, 1.00, 1.00),
("carpet", 0.25, 0.60, 0.90, 0.90),
("grass", 0.40, 0.40, 0.70, 0.75),
("gravel", 1.00, 0.20, 0.60, 0.60),
]
def __init__(
self,
depth_std_threshold: float = 0.02,
reflectance_window_size: int = 10,
hysteresis_count: int = 5,
control_rate_hz: float = 10.0,
):
self._depth_std_threshold = depth_std_threshold
self._hysteresis_count = hysteresis_count
self._depth_vars = deque(maxlen=reflectance_window_size)
self._reflectances = deque(maxlen=reflectance_window_size)
self._current_type = "unknown"
self._current_confidence = 0.0
self._candidate_type = "unknown"
self._candidate_count = 0
@property
def current_surface_type(self) -> str:
return self._current_type
@property
def current_confidence(self) -> float:
return self._current_confidence
def update(self, depth_variance: float, reflectance_mean: float) -> TerrainClassification:
"""Update classifier with sensor measurements."""
roughness = min(1.0, depth_variance / (self._depth_std_threshold + 1e-6))
reflectance = np.clip(reflectance_mean, 0.0, 1.0)
self._depth_vars.append(roughness)
self._reflectances.append(reflectance)
avg_roughness = np.mean(list(self._depth_vars)) if self._depth_vars else 0.0
avg_reflectance = np.mean(list(self._reflectances)) if self._reflectances else 0.5
candidate_type, candidate_confidence, speed_scale = self._classify(
avg_roughness, avg_reflectance
)
if candidate_type == self._candidate_type:
self._candidate_count += 1
else:
self._candidate_type = candidate_type
self._candidate_count = 1
if self._candidate_count >= self._hysteresis_count:
self._current_type = candidate_type
self._current_confidence = candidate_confidence
return TerrainClassification(
surface_type=self._current_type,
confidence=self._current_confidence,
roughness=avg_roughness,
reflectance=avg_reflectance,
speed_scale=speed_scale,
)
def reset(self) -> None:
"""Clear buffers and reset to unknown."""
self._depth_vars.clear()
self._reflectances.clear()
self._current_type = "unknown"
self._current_confidence = 0.0
self._candidate_type = "unknown"
self._candidate_count = 0
def _classify(self, roughness: float, reflectance: float) -> tuple[str, float, float]:
"""Classify terrain and compute confidence + speed scale."""
best_match = None
best_score = 0.0
best_speed = 1.0
for name, rough_max, refl_min, refl_max, speed in self._SURFACE_RULES:
rough_penalty = max(0.0, (roughness - rough_max) / rough_max) if roughness > rough_max else 0.0
refl_penalty = 0.0
if reflectance < refl_min:
refl_penalty = (refl_min - reflectance) / refl_min
elif reflectance > refl_max:
refl_penalty = (reflectance - refl_max) / (1.0 - refl_max)
score = 1.0 - (rough_penalty + refl_penalty) * 0.5
if score > best_score:
best_score = score
best_match = name
best_speed = speed
confidence = max(0.0, best_score)
return best_match or "unknown", confidence, best_speed
@staticmethod
def speed_scale_for_surface(surface_type: str) -> float:
"""Get speed scale for a surface type."""
for name, _, _, _, speed in TerrainClassifier._SURFACE_RULES:
if name == surface_type:
return speed
return 0.75

View File

@ -0,0 +1,4 @@
[develop]
script_dir=$base/lib/saltybot_terrain_classification
[install]
script_dir=$base/lib/saltybot_terrain_classification

View File

@ -0,0 +1,25 @@
#!/usr/bin/env python3
from setuptools import setup, find_packages
setup(
name="saltybot_terrain_classification",
version="0.1.0",
packages=find_packages(),
data_files=[
("share/ament_index/resource_index/packages",
["resource/saltybot_terrain_classification"]),
("share/saltybot_terrain_classification", ["package.xml"]),
],
install_requires=["setuptools"],
zip_safe=True,
maintainer="SaltyLab",
maintainer_email="seb@example.com",
description="Terrain classification using RealSense depth and RPLIDAR (Issue #469)",
license="MIT",
tests_require=["pytest"],
entry_points={
"console_scripts": [
"terrain_classification_node=saltybot_terrain_classification.terrain_classification_node:main",
],
},
)