PyRoboVision

Multi-object tracking (Kalman filter + Hungarian association) and trajectory prediction for robotics — pure Python (NumPy + SciPy) core, with an optional real MiDaS monocular depth backend.

What it is

PyRoboVision's real, tested core is a multi-object tracker: Kalman-filter state estimation plus Hungarian-algorithm association, with occlusion handling (tracks are predicted through detection gaps and re-associate on reacquisition) and constant-velocity / constant-acceleration trajectory prediction with uncertainty. Around it are supporting utilities: 3D perception (depth, 3D boxes, LiDAR, occupancy grids), rule-based behavior/intent classification, and minimal imitation-learning / behavior-cloning / safety-constraint building blocks. The core depends on NumPy and SciPy only; PyTorch is optional.

Features

• Multi-object tracking — MOTTracker, Track, Detection; birth/confirm/occlude/death lifecycle; IoU association via scipy.optimize.linear_sum_assignment.
• Occlusion handling — every track is predicted each frame; a confirmed track survives up to max_age missed frames and re-associates to the same track_id.
• Trajectory prediction — TrajectoryPredictor with "cv" (constant velocity) or "ca" (constant acceleration) models and uncertainty.
• Depth estimation — DepthEstimator(model="heuristic") (default, no extra deps) or model="midas" (real pretrained MiDaS_small via torch.hub, optional [depth] extra).
• 3D perception utilities — Box3DConverter (2D bbox + depth → 3D), LiDAR point-cloud filtering/clustering, BEV occupancy grids.
• Sensor fusion — Kalman-based IMU/GPS fusion with covariance (SensorFusionEngine).
• ONNX export — thin wrapper around torch.onnx.export (optional [onnx] extra).
• Learning helpers — imitation learning, behavior cloning, SafetyValidator constraint correction.

Who it's for / Use cases

• Adding tracking on top of your own detector — feed bounding boxes from YOLO/SAM/anything into MOTTracker; it handles identity persistence through occlusion.
• Forecasting from tracked positions — constant-velocity/acceleration prediction with uncertainty for downstream collision-avoidance logic.
• IMU/GPS fusion — a fused position estimate with covariance without hand-rolling a Kalman filter.
• Prototyping 3D perception — depth maps, 3D boxes, and occupancy grids from 2D detections and a calibrated camera.

What it is not

• No object detector. You bring your own detections (bounding boxes + confidence); none is bundled or wrapped.
• No GPU-accelerated inference pipeline. The tracking/prediction core is CPU NumPy/SciPy.
• No foundation-model integration (CLIP/SAM/Grounding DINO). An earlier version shipped a module that returned hardcoded fake results; it was removed, not fixed.
• MiDaS output is relative, not metric — correctly ordered near/far depth, not meters, unless you calibrate against a known reference.
• Not a certified or safety-verified autonomous-driving stack. SafetyValidator is a basic constraint checker.
• The default heuristic depth backend is a Sobel-edge placeholder, not a learned depth model.

Project status

Beta (Development Status 4). Current source version is 3.3.0; the latest PyPI release is 3.2.0. tracking/ and prediction/ are implemented and tested; supporting utilities (perception, fusion, learning) are real but minimal. 304 passed / 8 skipped in the local suite; the skips are optional torch/MiDaS paths that run when the [depth] extra is installed. black / mypy are not enforced in CI today. See ROADMAP_HONEST.md.

ByteTrack-style two-stage association

ByteTrackTracker keeps low-confidence detections for a second association pass instead of throwing them away, improving continuity through short occlusions without adding Ultralytics/AGPL code:

from pyrobovision import ByteTrackTracker

tracker = ByteTrackTracker(high_threshold=0.5, low_threshold=0.1)
tracks = tracker.update(detections)

Only high-confidence detections create new tracks; low-confidence detections may recover existing tracks. The implementation reuses the existing Kalman/Hungarian primitives and is unit-tested with synthetic boxes.

Bring your own detector (Supervision adapter)

PyRoboVision still does not ship a detector. It now accepts detections from Supervision (MIT), which normalizes outputs from MMDetection, YOLO-NAS, and other permissive detectors:

pip install "pyrobovision[supervision]"

from pyrobovision import SupervisionTrackerAdapter

adapter = SupervisionTrackerAdapter()
tracked = adapter.update(sv_detections)  # sv.Detections with tracker_id

detections_from_supervision() and tracks_to_supervision() are also exported for custom pipelines. Ultralytics remains unsupported because it is AGPL-3.0.

Alpamayo trajectory and reasoning adapter

Alpamayo (inference code Apache-2.0) predicts ego trajectories and Chain-of-Causation reasoning. The optional adapter normalizes its NumPy/tensor outputs without importing torch or alpamayo_r1:

from pyrobovision import AlpamayoTrajectoryAdapter

adapter = AlpamayoTrajectoryAdapter()  # 64 waypoints at 10 Hz
output = adapter.parse(pred_xyz, pred_rot, extra)  # upstream model outputs
best, metrics = adapter.select_best(output, ego_future_xyz)

output.trajectories exposes (x, y, z) waypoints, optional rotations, and per-sample reasoning; positions_2d() drops z for 2D consumers. Alpamayo 1 weights are research/non-commercial and inference requires an NVIDIA GPU, so model execution stays outside this package.

Installation

pip install pyrobovision

# real MiDaS depth estimation (installs PyTorch; one-time model download)
pip install "pyrobovision[depth]"

# ONNX model-export utilities
pip install "pyrobovision[onnx]"

Requires Python 3.10+. From source:

git clone https://github.com/Mullassery/PyRoboVision.git
cd PyRoboVision
pip install -e ".[dev]"

Quick start

import numpy as np
from pyrobovision.tracking.mot import MOTTracker, Detection

tracker = MOTTracker(max_age=30, min_hits=3)

# Detections come from *your* detector — each is a bounding box + confidence.
detections = [Detection(bbox=np.array([100, 100, 140, 180]), confidence=0.92)]

for track in tracker.update(detections):
    print(track.track_id, track.get_position(), track.get_velocity())

Usage

Tracking plus trajectory prediction:

import numpy as np
from pyrobovision.tracking.mot import MOTTracker, Detection
from pyrobovision.prediction.trajectory import TrajectoryPredictor

tracker = MOTTracker(max_age=30, min_hits=3)
positions = []
for x in (100, 110, 120, 130):
    tracker.update([Detection(bbox=np.array([x, 100, x + 40, 180]), confidence=0.9)])
    positions.append(tracker.get_track_by_id(1).get_position())

predictor = TrajectoryPredictor(model="cv")  # "cv" or "ca"
print(predictor.predict_trajectory(np.array(positions), horizon=5))

Depth + 3D boxes + occupancy (default heuristic depth; swap in model="midas" for real inference):

import numpy as np
from pyrobovision.perception.depth import DepthEstimator
from pyrobovision.perception.bbox_3d import Box3DConverter
from pyrobovision.perception.occupancy import OccupancyGridBuilder

estimator = DepthEstimator(model="heuristic")
estimator.set_calibration(fx=500, fy=500, cx=320, cy=240)
depth = estimator.estimate_depth(np.random.randint(0, 255, (480, 640, 3), dtype=np.uint8))

box = Box3DConverter().from_2d_bbox_and_depth(
    bbox_2d=np.array([100, 100, 200, 200]), depth_map=depth.data, fx=500, fy=500, cx=320, cy=240)
grid = OccupancyGridBuilder(grid_size=(100, 100), resolution=0.1).from_3d_bboxes([box])
print(len(grid.get_occupied_cells()))

IMU/GPS fusion and safety checks:

from pyrobovision.fusion.sensor_fusion import SensorFusionEngine, IMUData, GPSData
from pyrobovision.learning.safety import SafetyValidator

engine = SensorFusionEngine(origin_lat=37.7749, origin_lon=-122.4194)
engine.update_imu(IMUData(timestamp=0.0, accelerometer=[0, 0, 9.8], gyroscope=[0, 0, 0]))
state = engine.update_gps(GPSData(timestamp=0.1, latitude=37.7749, longitude=-122.4194,
                                  altitude=10.0, speed=5.0, heading=90.0, accuracy=2.0))

safe = SafetyValidator(max_acceleration=5.0, max_steering=45.0).correct_action(
    proposed_action=[6.5, 50.0], context={"speed": 15.0})  # clamped to limits

Documentation

• examples/ — basic_tracking.py, depth_estimation.py, real_detector_tracking.py
• CONTRIBUTING.md — development setup and guidelines
• ROADMAP_HONEST.md — done / not done / next, and technical debt
• docs/BENCHMARKS.md — benchmark status and how to run your own
• docs/CODE_OF_CONDUCT.md — community guidelines
• CHANGELOG.md — notable changes
• SECURITY.md — vulnerability reporting

Known issues and deferred work

Blocked / failed:

• Detector/GPU integration; otherwise freeze.

Rejected / deferred:

• Foundation models/CLIP/SAM.

Full shipped/blocked/rejected detail: QUEUE_STATUS.md.

Contributing

Contributions welcome — see CONTRIBUTING.md. Security issues: SECURITY.md.

Part of the robotics & autonomy family

PyRoboVision is one of six complementary, composable tools in the Mullassery robotics & autonomy family. Together they form a single stack: perceive → map → simulate → learn → debug → coordinate the fleet.

Each works standalone, but they're complementary by design. Browse all projects →

License

Apache License 2.0 — see LICENSE.