Computer Vision · Embedded Linux · Python

RPi5
Object
Orientation
Tracker

Real-time orientation measurement for objects in a live camera feed, running on a Raspberry Pi 5 with 1GB RAM - no GPU, no cloud. Haar cascade detection, Canny edge + Hough transform for angle, sliding window filter for stable output, CSV logging. Extended from the BiViP image processing lab at FAU Erlangen-Nürnberg.

Hardware Raspberry Pi 5

Speed 10 fps

Resolution 640 × 480

Stack Python · OpenCV

Frame budget 100 ms

View on GitHub → How it works

01Camera inputPicamera2 · 640×480 · XRGB8888

02PreprocessGaussian blur 5×5 · grayscale

03DetectionHaar cascade · scaleFactor 1.1

04ROI crop±200px margin · boundary check

05Canny edgesTlow=50 · Thigh=150 · apt=3

06Hough linesSelect line closest to 90°

07Angletheta → degrees · 0–180 range

08Filterdeque window=5 · clear on miss

09CSV logtimestamp · raw · filtered · count

10DisplayBounding box · angle overlay

—

Run it

# clone
git clone https://github.com/VaibhavAher100/rpi-object-orientation-tracker
cd rpi-object-orientation-tracker

# laptop — no Pi required for testing
pip install -r requirements.txt
python src/detector.py --no-picamera

# on the Pi
pip install -r requirements-rpi.txt
python src/detector.py

# plot the angle trace after a run
python src/visualize_results.py

# tune the detector
python src/detector.py --window 8 --cascade classifiers/pen_vertical_classifier.xml
  

—

Pipeline detail

Each stage feeds the next. The design choice that matters: in stage 06, the code selects the Hough line closest to vertical (90°) rather than blindly taking the first line returned. Blindly taking lines[0] gave unstable angle readings when multiple lines were detected.

Camera input

Picamera2 captures 640×480 in XRGB8888. --no-picamera flag swaps in cv2.VideoCapture for laptop testing.

Preprocess

Gaussian blur (5×5, σ=0) then grayscale. Reduces noise before the cascade fires.

Detection

detectMultiScale — scaleFactor=1.1, minNeighbors=5, minSize=(40,40). Returns a list of bounding boxes.

ROI crop

Expand the bounding box by ±200px / ±100px. Falls back to ±40px / ±20px when near the frame edge.

Canny edges

Applied to the cropped grayscale region. Tlow=50, Thigh=150, apertureSize=3.

Hough lines

cv2.HoughLines on the edge image. Selects the line with theta closest to π/2 (vertical). Returns -1.0 if nothing found.

Angle

theta × 180/π. HoughLines returns theta in [0, π) so the result is already in [0, 180) degrees.

Moving average

deque(maxlen=5). Window clears when object count drops to zero — prevents stale angle bleeding into the next detection.

CSV log

One row per frame: unix timestamp, raw_angle, filtered_angle, object_count. Header written on first call.

Display

Green bounding box. Angle printed top-left in red. cv2.imshow — press Q to quit.

—

Limitations

The worst-case detection sample (three false positives) is committed to results/. These limitations are not hidden.

False +

Cable edge fires the classifier. A vertical cable on the right of the frame has a similar edge profile to a vertical pen. Raising minNeighbors reduces it but also misses weaker real detections.

Laptop

Classifier was not trained for general objects. When testing on a person video on laptop, it fires occasionally on jacket and collar edges, returning 0.0 angle. The cascade knows vertical pens, not everything.

Multi

Only the first detection gets an angle. object_count reflects all bounding boxes but raw_angle and filtered_angle come from the first box only. Tracking angle per object ID was out of scope.

Light

Low contrast breaks detection. The cascade needs reasonable lighting contrast. Tuning scaleFactor down helps sensitivity but slows the pipeline.

RPi5ObjectOrientationTracker

Run it

Pipeline detail

Limitations

RPi5
Object
Orientation
Tracker