Robert Crowe

Created a computer vision pipeline capable of detecting vehicles in video stream in real-time:

* Performed a Histogram of Oriented Gradients (HOG) feature extraction on a labeled training set of images and train a Linear SVM classifier

* Applied color transform and added binned color features, as well as histograms of color, to HOG feature vector

* Implemented a sliding-window technique and use your trained classifier to search for vehicles in images

* Estimated a… Show more

Created a computer vision pipeline capable of detecting vehicles in video stream in real-time:

* Performed a Histogram of Oriented Gradients (HOG) feature extraction on a labeled training set of images and train a Linear SVM classifier

* Applied color transform and added binned color features, as well as histograms of color, to HOG feature vector

* Implemented a sliding-window technique and use your trained classifier to search for vehicles in images

* Estimated a bounding box for vehicles detected Show less

Implemented an advanced computer vision approach to identifying and tracking the lane boundaries and the vehicle position using data from a forward facing camera:

* Computed the camera calibration matrix and distortion coefficients given a set of chessboard images.
* Applied distortion correction to raw images for correct for lens distortion
* Used color transforms, gradients, etc., to create a thresholded binary image.
* Appled a perspective transform to rectify binary image… Show more

Implemented an advanced computer vision approach to identifying and tracking the lane boundaries and the vehicle position using data from a forward facing camera:

* Computed the camera calibration matrix and distortion coefficients given a set of chessboard images.
* Applied distortion correction to raw images for correct for lens distortion
* Used color transforms, gradients, etc., to create a thresholded binary image.
* Appled a perspective transform to rectify binary image ("birdseye view").
* Detected lane pixels and fit to find the lane boundary.
* Determined the curvature of the lane and vehicle position with respect to center.
* Warped the detected lane boundaries back onto the original image.
* Outputted a visual display of the lane boundaries and numerical estimation of lane curvature and vehicle position, overlaid on the camera video Show less

* Sole developer
* Trained a model to steer a car in a simulator, using the steering data generated as a human drives the same car as the training data. It is therefore a regression problem, since steering angles are continuous values.
* Successfully trained the model to steer the car correctly 100% of the time

* Sole developer
* This is an implementation in Tensorflow of a convolutional network classifier for the German Traffic Sign dataset. http://benchmark.ini.rub.de/?section=gtsrb&subsection=dataset
* The architecture is based on a paper by Pierre Sermanet and Yann LeCun: http://yann.lecun.com/exdb/publis/pdf/sermanet-ijcnn-11.pdf
* Achieved 80.95% accuracy on a challenging dataset

* Sole developer
* An implementation of AlexNet in Tensorflow to classify Google Street View House Numbers. AlexNet is a Convolutional Neural Network and classifier that established CNNs as the leading architecture for image processing.
* Achieved 93.56% accuracy

* Sole developer
* An implementation of reinforcement learning using Q-learning to teach a smart cab to drive. Implemented in Python using Scikit-Learn (sklearn).
* Starting with a model that makes random choices and receives rewards and penalties based on the scenario rules, the model is analysed and improved as Q-learning is implemented, the state-action pairs are defined, and hyperparameters are adjusted. The implementation goes through four different versions of incremental… Show more

* Sole developer
* An implementation of reinforcement learning using Q-learning to teach a smart cab to drive. Implemented in Python using Scikit-Learn (sklearn).
* Starting with a model that makes random choices and receives rewards and penalties based on the scenario rules, the model is analysed and improved as Q-learning is implemented, the state-action pairs are defined, and hyperparameters are adjusted. The implementation goes through four different versions of incremental refinement.
* Acheived 99% success rate Show less

Developed system to predict passenger wait times based on sensor data. Currently in production at Detroit International Airport:

* Field tested 3 different sensor systems and data to establish cost/value

* Developed and trained feed-forward neural network using Tensorflow

* Developed production system to perform inference using trained model using Tensorflow Serving

* Developed web-based management and reporting console