???:
- Real time driver fatigue detection
- Dashcam/laptop
- Algorithms:
- Gaussian blur
- Histogram of oriented gradients (HOG)
- SVM
- Percentage eyelid closure (PERCLOS); eye aspect ratio (EAR), mouth aspect ratio (MAR)
???:
- Wildfire hotspot detection
- Smouldering makes detection difficult
- Heavy duty water carrying drone
- Identify, rank, then approximate distance to hotspot
- Drone camera
- Algorithms:
- Segmentation: threshold, contour, centroids
- Morphology
- Distance approximation (camera calibration)
???:
- Same as above
- Algorithms:
- Gaussian
- Binary threshold
- Morphology: opening/erosion/dilation
- Contour detection
- Find largest centroids
- Lucas-Kanade optical flow: is contour expanding
- Stereo imaging for distance approximation?
???:
- Sign language teaching assistant
- Real time feedback
- Only RGB camera
- Algorithms:
- Sharpening kernel
- Convolutional pose machine
- Thresholding
- Hand keypoint detection at some point
- 0.6 Hz
???:
- Track moving objects in robot soccer
- Algorithms:
- Camera calibration
- Background subtraction
- Circular Hough transform + unscented Kalman filter for ball tracking
- CNN object detection for robot detection
???:
- Faster human detection
- Want low power, fast systems
- CNN-based object detection
- Algorithms:
- YOLOv5
- Kalman filter
- Tracker fit model
- 0:11:54
- Run Kalman prediction
- Run object detection
- Are they close enough (Euclidean)?
- No - create new tracker
- Yes - update Kalman with new location
- Multiple objects within max distance: pick closest one
- Low CNN depth, Kalman to remove flicker/missed detection in frames
???:
- Gorilla head tracking
- Algorithms:
- Greyscale frame difference: subtraction
- Binary thresholding
- Median blur
- Morphology: open
- Find centroid
- Moment
- Single person only
???:
- Wildfire
- Algorithms:
- Mean greyscaling
- Gaussian
- To-zero thresholding
- x > threshold? x: 0
- Circular kernel (blurring?)
- Pick brightest pixel
- Blackout pixels around the selected pixel; repeat to get next brightest pixel
- Guess location by assuming ground is flat: know drone position, camera angle
???:
- FPS enemy detection
- Algorithms:
- YOLOv5 to get bounding box
- For outlining:
- Gaussian
- Morphology: close
- Morphology: gradient (difference of dilation and erosion)
- Thresholding: Otsu
???:
- Rehab to free throw detection (basketball)
- Detectron2 pose detection: pyramid network + Mask R-CNN
- Nothing else?
???:
- Chess board framing, move detection
- Lighting can change
- HSV color masking to detect chess board red/blue pieces
- Algorithms:
- Board framing:
- HSV color masking to remove background
- Morphology: open/close
- Contour finding: find square
- Chess grid:
- Canny edge
- Hough line transform
- Homography matrix
- Chess piece:
- HSV Color masking
- Board framing:
???:
- Construction panel quality control
- Detect dimensions, end tolerances
- Algorithms:
- Processing:
- Greyscale
- Blur
- Mask
- Canny edge detection
- Morphology: erosion/dilation
- Calibration:
- Find sticker - pixels to mm
- Find panel:
- Draw contour
- Template match: adapted normalized cross-correlation
- Processing:
???:
- Detect/localize pine tree yellow catkins
- Real time
- Algorithms:
- YOLOv5 catkin detection
- MeanShift segmentation of depth image
- Morphology: opening/closing
- Contour detection Suzuki’s algorithm
- Shape matching: Hu moment invariants
???:
- Identify cut logs
- YOLACT with custom dataset for instance segmentation
???:
- Othello piece detection
- Algorithms:
- Gaussian
- Color thresholding
- Douglas-Peucker algorithm
- Image warping: homography matrix
???:
- Quick-time event detection in Detroit: Become Human
- QTE: fast button/gesture prompt
- Algorithms:
- Hough circle after greyscale, median blue
- Text detection:
- Tesseract OCR
- Frame crop
- Bicubic interpolation
- Gaussian
- Morphology: Erosion
- Grayscale
- Thresholding
- Harris corner detector to detect symbols
- Not sure how symbol type is identified
???:
- Rubiks cube tracking, photo -> model of state
- Stickerless cubes - edges not well defined (no black border)
- Algorithms:
- Harr cascade classifier to detect cube - generate bounding box
- Split into 3x3 grid
- Median blur
- Sample center
- Color thresholding to classify color
- Use LAB color space - less sensitive to luminance
- Also used: Canny edge detection, Suzuki85, Douglas-Peucker
- Contour detection: OpenCV, Suzuiki85 https://doi.org/10.1016/0734-189X(85)90016-7
???:
- Integral calculator
- Algorithms:
- Greyscale
- Gaussian
- Binary threshold
- Morphology: erode/dilate
- Find contours, sorting by x value
- Tessaract OCR
- Convert to string
- Compute integral
- Fails with handwriting, bad with printed, good with screenshots
???:
- Squash tracking in 3D space
- Position of squash ball
- Algorithms:
- Subtraction from clean plate (different from background modeling?)
- Morphology: erosion for noise removal
- Use output as mask
- Contour detection: group
- Filter by size, shape
- Triangulate position using two different cameras
- Triangulation with two cameras
???:
- Automatic projector keystone calibration
- Algorithms:
- Canny edge detection
- Hough to detect image edges
- Ramer-Douglas-Peucker
- Homography transform
???:
- Pose estimation: squat depth
- Algorithms:
- Segmentation mask to separate athlete from background
- MediaPipe pose
- Use keypoints to estimate squat depth
???:
- Virtual paper piano
- Printout of keyboard
- Algorithms:
- Keyboard segmentation:
- Canny edge detection
- Merged Hough line transform
- Perspective transform
- Linear segmentation
- Fingertip tracking and touch detection
- MediaPipe
- Transform to perspective space
- Keyboard segmentation:
- Only works with a single finger
???:
- Robocup object detection - identify weight
- Algorithms:
- Stereo calibration (estimate parameters)
- Gaussian
- Template matching: normalized correlation coefficient
- Triangulation to estimate depth
???:
- Bicycle detection and camera tracking (with gimbal)
- Algorithms:
- YOLO v5 to detect bike
- Pass bounding box to CSRT tracking algorithm (Channel and Spatial Reliability Tracking)
???:
- Stylus input with CV
- Digital sketchpad
- Algorithms:
- Canny
- Hough
- Shi-Tomasi corner detection
- Morphology: opening
- Stylus has colored tab which gets exposed when pressure is applied to the tip
- HSV filter
???:
- Predictive motion
- Algorithms:
- Gaussian blur
- Morphology dilation/opening
- Adaptive thresholding
- Hough circle transform
???:
- Patient rehabilitation monitor; track pose when doing exercises
- Hardcoded reference angles
- Algorithms:
- CNN
- Joint angle calculation, comparison thresholding
- FSM for multistage verification
42:00
???:
- Melanoma detection using smartphone camera
- Identify moles, return measure for border irregularity and color variance
- Algorithms:
- Increase image contrast
- Gaussian
- Greyscale
- Morphology: closing
- Adaptive binary thresholding
- Suzuki-Abe contour
- Fitzgibbon ellipse fitting
???:
- Nuclei segmentation in breast cancer tissue images
- Algorithms:
- Segmentation with convolutional autoencoder (U-net)
- Thresholding
- Morphology
- Watershed
???:
- AR sudoku solver
- Overlay solution on paper
- Algorithms:
- Pre-processsing:
- Adaptive thresholding
- Morphology: open/dilate
- Grey scale
- Hough to detect horizontal/vertical lines:
- Remove the lines to be left with numbers only
- Perspective transform to warp and crop image
- CNN for digit classification
- Solve sudoku using backtracking, render added text, then unwarp to overlay on top of input image
- Pre-processsing:
???:
- Tumor/tissue detection: generate tissue and bulk region masks
- Algorithms:
- Morphology: erosion/dilation
- Contour detection
- Color thresholding
- Median blurring
???:
- Drone detection in airports, differentiate between drone and birds
- Algorithms:
- Double difference w/ sharpen kernel to detect difference
- Contour extraction (using morphology - close/dilate)
- Intersection over union tracker
- Moving average filter
- Fast fourier transform to extract power spectral density
- K-means/linear classifier
???:
- Automatic exposure control for robot navigation
- Algorithms:
- Sobel gradient filter
- Gradient magnitude
- Soft percentile derivative: weighted sum of difference between two frames
- Slice image into equal sections: take median of soft percentile derivatives
- No way of automatically calculating number of slices
???:
- Punching technique stats
- Track fist velocity, acceleration, elbow angle, fist angle
- Algorithms:
- Detectron2 retinanet for fist top/side detection
- Mediapipe pose
- Non-maximum suppression
???:
- Football penalty ball tracking
- Camera head-on with goal, located behind player
- Detect goal posts and ball
- Used yellow ball: HSV color range used to detect ball
- Algorithms:
- Gaussian
- Morphology: open, close, erode, dilate
- Thresholding
- Contours
???:
- Joint tracking: give likeness score computed from joint angles
- D435 camera
- ~0.4x real time speed
- Algorithms:
- Region-based CNN (R-CNN): instance segmentation (detectron 2)
- Keypoint/pose detection?
- Dynamic time warping
- Linear interpolation (for dealing with varying framerates)
- Newton’s method (to align video sequences)
- Region-based CNN (R-CNN): instance segmentation (detectron 2)
???:
- Scrabble board detection
- Continuing code from previous student project
- Algorithms:
- Greyscale
- Adaptive thresholding
- Detect board contours
- Crop image to contain only board?
- HSV masking: filter out undesired colors (i.e. non letter-tiles)
- Morphology: erosion/dilation to remove noise
- Detect ‘maximally stable extremal regions’
- Tesseract OCR
???:
- Rugby ball detection
- Static camera, players in frame
- Difficulty: rugby balls not circular
- Algorithms:
- Gaussian background subtraction
- Median filter
- Morphology: close
- Contour detection
- Filter:
- By area
- Compactness degree: contour area divided by area of best-fit ellipse
- By ellipse aspect ratio
???:
- Pet detection using dominant color
- Algorithms:
- Mask-R CNN instance segmentation to detect pets in frame
- K-means clustering to determine dominant pet colors
- Image thresholding to remove image background?
???:
- Handwritten digit recognition
- Non-linear transform of input features into higher dimensional space so that features are linearly separable
- Cheaper than conventional deep learning
- Nanowire network: randomly? scattered nanowires combine to form network with complex topological structures: junctions between wires act as a form of non-linearity and memory
- Create 3D simulation of wires: do not assume they are 1D lines
- Input: voltage into input electrodes (1 out of 4 edges of a square)
- Output: current from output electrodes (remaining 3 edges of the square)
- Train only the output layer through regression to make 10 classifiers
- Algorithms:
- Linear classifier:
- Moore-Penrose pseudoinverse
- Singular value decomposition (SVD)
- Single layer neural network using Tensorflow
- Softmax activation
- Catagorical cross-entropy
- Adam optimizer
- Nanowires deposition: Euler rotations
- Junctions model using Stormer-Verlet integration
- Modified nodal analysis to solve Kirchoff’s circuit laws
- Linear classifier:
???:
- Blackjack simulation: detect card rank and suit
- Top-down images of cards
- Algorithms:
- Greyscale
- Gaussian
- Thresholding
- Contours: detect cards
- Morphology, close: use closing to merge close cards into a single group to detect hands (i.e. based on distance)
- Image differencing: match corner to preset images to detect rank/suit
???:
- Motion detection for raster graphics editor
- i.e. draw using hand gestures
- Webcam facing user
- Algorithms:
- Image rectification
- Classification/localization CNN to crop to hand
- Instance segmentation CNN to detect keypoints
- Use relative distance between keypoints as gesture/controls
???:
- Real-time face replacement
- https://learnopencv.com/face-swap-using-opencv-c-python
- Algorithms:
- Face detection with
dliblibrary- HOG + linear SVM face detector?
- Face alignment: convex hull
- Delaunay triangulation
- Texture mapping by using affine warp to map triangles between the two images
- OpenCV Seamless cloning
- Face detection with
???:
- Blood splatter analysis
- Algorithms:
- Pre-processing:
- Thresholding
- Dilation
- Resizing
- Pre-processing:
- CNN training: ResNet-50
???:
- Paper piano
- Use built-in laptop camera (paper on top of keyboard/trackpad?)
- One finger only
- Algorithms:
- Adaptive thresholding (with Gaussian mean)
- Morphology: opening for noise reduction
- Ramer-Douglas-Peucker for contour detection
- Finger detection with color thresholding
???:
- Real time number input for timer control using static hand gestures
- Algorithms:
- MediaPipe hands
- Gesture classification:
- Feature-angle thresholding
- Support vector machine (SVM)
- Debouncing:
- https://ieeexplore.ieee.org/abstract/document/8868766
- Time/frame-delay debouncing
???:
- Risk parameter of ALC injury risk
- Record jumping video
- Depth camera
- Detect 4 risk factors
- Algorithms:
- Detectron2 keypoint detection
???:
- Hand gesture controlled calculator
- Use vector of each finger as an input bit
- Thumbs were a special case
- Algorithms:
- Hand keypoint detection
- Kalman filter predictions
???:
- Butterfly/moth classification
- Algorithms:
- Instance segmentation
- Non-maximum supression
- Detectron2 to crop
- resnet18 to classify
???:
- Blood vessel extraction from image
- Algorithms:
- Binary thresholding
- Morphology: open/close/erosion
- CLAHE algorithm to increase contrast
- Illumination equalization
- Gaussian
- Otsu’s threshold: vessel segmentation
???:
- Hand gesture recognition for sign language w/ smartphone camera
- Algorithms:
- Histogram
- Histogram backprojection
- Morphology: closing
- Inception v3 neural network
???:
- SLAM with monocular video
- Algorithms:
- Previous and current frame input into ORB keypoint detection
- FLANN keypoint matching
- Lowe’s ratio match pruning
- RANSAC 5-point: generate essential matrix
- Velocity information for pose recovery
- Filter spurious transforms
- Increment by transformations
???:
- Cow teat detection
- Camera under cow
- Algorithms:
- Blob detection
- Thermal filter - coldest = teat
- Filter by circularity, aspect ratio
- Morphology: erosion/dilation
- Repeat with different morphology parameters until four teats in valid shape detected
- Blob detection
???:
- Visual cue for call detection for pass gesture in basketball?
- Detect gesture from player so that robot shoots ball?
- Algorithms:
- Body pose estimation with Detectron2
- Identify largest skeleton in frame
- Locate wrist
- Extract subframes around wrist (proportional to size of skeleton)
- Mediapipe hand pose
- Measure distances between certain keypoints to detect gesture
- Threshold
- Pass threshold pass/fail to FSM
- Body pose estimation with Detectron2
???:
- Sudoku detection/solving
- Algorithms:
- Gaussian
- Adaptive Gaussian thresholding
- Contour filling: identify contours to determine board outline
- Green’s theorem
- Warp image to board shape
- OCR to identify numbers
- Solve board using backtracking
???:
- Real time NZ sign language detection
- Algorithms:
- Pre-process CNN input:
- Binary thresholding
- Issues: lighting, busy backgrounds
- Canny edge detection
- Binary thresholding
- CNN
- Pre-process CNN input:
???:
- Emotion detection of audiences (multiple people)
- Deepface for facial recognition
- 9 layer NN
- Algorithms:
- Haar cascade
- Face frontalization
- 2D/3D alignment
???:
- Speed limit recognition of NZ street signs
- Detect, then read
- Detection:
- Color space transform to emphasize red (CIELUV)
- Gaussian
- Circle hough transform
- Text
- Crop to isolate text
- Filter
- Threshold: Otsu’s method
- Tesseract OCR
???:
- Real time dart scoring
- Identify dart throw events
- Algorithms:
- Create foreground mask?
- Board detection:
- Canny edge
- Hough transform
- Morphology
- Contouring
- Dart detection:
- Background subtraction
- Morphology
- Contouring
- Keypoint detection with YOLOv4-tiny
???:
- Stereo imaging to find distance/orientation of plane
- Want to require low overlap
- Crop images (Right edge/left edge: overlap)
- Create disparity map
- Normalize values
- Convert to depth
- Average values
- Reproject to 3D
???:
- Emotion recognition through facial expression
- Algorithms:
- Haar cascades face recognition
- Facial action coding system
- Augmentation for classifier training: brightness, rotation, shift
- Classifier: CNN + softmax
???:
- Measuring body dimensions with depth camera
- Use 3D points to get dimensions
- Didn’t get to algorithms – out of time
???:
- Music sheet reader for the visually impaired
- Use eye tracking to zoom in and pan into the sheet
- Prior research: automatic page turning but no zoom
- Algorithms:
- Template matching
- Gaussian
- Region of interest
- Gradient orientation pattern (eye tracking)
???:
- Mobile pool ball detection/identification
- Track position of all balls and identify for scoring
- Prior research: permanent setups
- Try use hand-held video
- Algorithms:
- Gaussian
- Hough circle transform to mask ball
- Get average color of mask
- Use color to identify ball score
- Low success rate: 38%
???:
- Inventory stocktaker
- Repeated patterns: instance segmentation?
- One phone camera, using flash to try control lighting
- Detects repeated vertical lines to count items
- Algorithms:
- Countouring
- Hough lines
- Angle filter
- Morphology: erode/dilate
- Hough lines
???:
- Beer pong score keeper
- Algorithms:
- Gaussian
- Ball detection: color mask + blob detection
- Cups: hough circle transform
- Issues:
- Top-down only
- Lighting/shadows
- Hardcoded size? Distance/lens combination fixed
???:
- Apple detection (for packaging robot)
- Create own dataset
- Algorithms:
- To remove background:
- HSV masking
- Morphology
- To remove background:
- Segmentation labeling for training data
- Mask R-CNN (detectron2)
???:
- Face recognition/tracking on multiple subjects without prior training
- Real time
- Store embedding of faces, store vectors and compare (HOG)
- Algorithms:
- Single shot multibox detector (resnet base)
- Ensemble of regression trees
- Embedding creation: ResNet
- Recognition: euclidean distance + linear embedding search
???:
- Document supermarket receipts
- Product names, cost, total cost
- Use scanner rather than smartphone camera
- Algorithms:
- Adaptive binaraization: Otsu’s method
- Morphology: erosion
- Tesseract OCR
???:
- Antarctic snowstorm classification with CNN
- ResNet-18 model
- Manually label dataset, then artificially grow
- Augment dataset: horizontal flip, crop, resize
- Classification:
- Random crops from image: classify each and combine
- Algorithms:
- Cross entropy loss function
- Stochastic gradient descent
- Ensemble method to combine
???:
- Cricket ball shot tracker
- Algorithms:
- Differencing
- Thresholding
- Median blur
- Dilation
- Contour detection
???:
- Guitar string picking detection
- Algorithms:
- Fretboard/frets/string detection:
- Canny edge
- Hough line
- Fretboard/frets/string detection:
- Pick location: user identifies initial position for template matching, then tracked
- Detecting picks: after velocity goes above threshold, find rapid deceleration
???:
- Hand gestures to control interactive display
- MediaPipe
- ML to recognize gestures using keypoints
- Limited input speed
???:
- Dartboard segmentation
- Single camera
- Algorithms:
- HSV color mask
- Morphology: open/close
- Flood filling
- Edge segmentation:
- Canny edge
- Hough line
- Centroid calculation
- Point multipliers (thin green/red regions): HSV + morphology
- Scoring regions: flood filling + bitwise operations
- Wedges: canny + hough line (with centroid to determine board center)
???:
- Cricket batting shot classification with pose estimation
- Algorithms:
- Gaussian
- Canny edge detection
- Detectron2 pose estimation
- SKLearn
???:
- Dirty dishes on kitchen bench
- Prior: Hough circle, Mask-RCNN
- Solution: Mask-RCNN with COCO dataset
- Plates: transfer learning using resnet?
- Motion detection to reduce false positives
- Gaussian
- Background segmentation
???:
- Darts scoreboard identification
- Blur: gaussian blur. HSV
- Dilate/erosion
- TODO
- Mask generation: thresholding
- Regionsegmentation: Hough lines, canny edge
- Motion detection - double differencing
- Flood fill
- Double differencing
- Triangle interior
???:
- Real time mask detection
- OpenCV/numpy face detection
- Keras/tensorflow face mask
- MobileNetV2
???:
- Manufacturing defect detectoin of surgical reamers
- Otsu thresholding
- Hough transform
- Morphology close
- Topographic TODO
- TODO 0:42
???:
- Robocup object detection, SLAM
- Identify cup weights
- Real time area mapping
- Haar cascade classifier
- ORB feature detection
- TODO
- RANSAC
- ORB-SLAM - didn’t work due to low color variance
???:
- Wheelchair docking assistant
- Prevent damage to desk/chair, injuries
- Slow down when close to the desk
- Object detection: instance segmentation (mask R-CNN)
- Distance calculation: stereo camera
???:
- Low cost stereo with two webcams
- Focus on speed: real time
- Two cameras taped together in a box
- Camera calibration
- Stereo rectification
- Disparity calculation
???:
- Determining queue times with CV
- Greyscale
- Detct faces
- Front-on faces
- Calcluate bounding box
- Centroid
- TODO
- Haar cascade classifier
- Masking
???:
- Measuring growth rates
- Crop/rotate.
- Blur, hsv theshold, erode/dilate
- Create contours of plants from mask
- Exclude contours outside of region of interest
- Find size of contours
- As leaves get closer to camera, visual size increases
???:
- Vehicle lane positioning with semantic segmentation
- If no road markings
- DeepLavV3 semantic segmenation
- Extract road surface, denoise road surface mask
- median blur
- Color thresold
- Morphology
- Canny edge
- TODO
- 0:53
???:
- Climbing tutor: track 3D poses
- Structure from motion
- Perspective n-point
- Locate keypoints: MediaPipe
- Triangulate
- Bundle adjustment: Least squares solver
???:
- Interactive musical webcam
- Volume/pitch depends on hand position
- Single shot detector
- Feature mapping
- Object prediction: CNN
- Object detection: TODO
- TODO
???:
- Real time background subtraction on people with HMD (AR)
- No green screen, single camera, no depth
- YOLACT for instance segmentation
- Mask
- Contour detection
- Thresholding
- TODO