Raymond Xu

Example of it in action

Here's a link to the extension in the chrome store. Check it out!

Sample vehicle detection and orientation prediction

Training infrared image detection or segmentation systems involves heavy data augmentation, and a balance of data diversity between different times of day, different outdoor temperatures, and sunshine levels.

Project Goal

Currently, 3D image registration is done with an optimization process, which is slow. Because the database of brain MRI scans is so large, the transformations can be memorized by a machine given the shape of a given brain. The goal is to speed this process up with a convolutional nerual networks so the process becomes one step instead of iterative, speeding up the process.

I designed a model to do this 3D registration with siamese convolutional layers and a spatial transformer network, using MSE as loss. The model worked well - it reduces the MSE between the input brains on the validation data set by 45%. Here is an architecture model outline:

Here is a demonstration of the feature maps and output from two input brains:

Here we can see the registration at work; it can register 3D brains with an average MSE difference between the registered brain and the source reduced by 45% compared to the source and the target originally. In this pair of brains, a slice of the brains are shown to show the successful registration.

ECE 3140: Embedded Systems

The final project here was using the FRDM kinetis k64f to light up an LED display. The difficulty here involved programming the board to be able to transmit i2c signals.

ECE 3400: Robot Competition

In this class, a group of 8 students and I had a project of making a self navigating robot that could navigate a maze like that shown in the video below. The robot was able to send a radio signal of its current location as well as the location of an infrared emitter located on the maze.

ECE 4250: Super Resolution Imaging

For this project, our group took multiple images of my room, calibrated them, added them, and averaged them to obtain an image with more information than any single image. This was supported with some image SNR metrics. Here is a link to the presentation for more information.

ECE 4760: Microcontrollers

The recording studio was programmed on the pic32 and allowed users to play the guitar, bass, drumset, or use their voice to make looped recordings, with additional features like deleting recordings or layering instruments on top of one another during recordings.

ECE 5725: Embedded Operating Systems

The dancing robot was a Raspberry Pi that listened to sounds with the microphone and performed beat detection in order to dance with the beat.

My work with the company was entirely building and perfecting a single model to classify radar emitters; even when those radars emitted signals that were different frequencies, frequency patterns, time lengths, and bandwidths.

Architecture Description

The architecture involved taking FFT's of the data, doing a couple convolutional filters on groups of them, and then inputting each FFT into a long short-term memory network (LSTM), one per time step. The LSTM states then learned how to interpret the signals. The output of the the LSTM is trained with an autoencoder, with the loss being the comparison with the original signal. Then, then network is trained by attaching the LSTM to several dense neural network layers. The network was able to achieve >95% test accuracy on the emitter classification.

This is the t-distributed stochastic neighbor embedding (tsne) viewing of the training data for the encodings of the dozen radar emitter classes. The data looks relatively separable.

This is the tsne of the test data. The encoder does not overfit, as the two look very similar.

Automated driver monitoring

Automated USPS driver monitoring for illegal/reckless driving using noisy GPS data. The program was designed to detect left turns and U-turns in the routes and map out their locations.

Cognitive radar classification

Researched and wrote an image/signal processing program to recognize enemy radar chirps in electronic warfare. The program receives a radar signal, transforms it into a spectrogram image, then filters that image and applies computer vision techniques to create recognizable features. A machine learning neural net then uses the processed image to classify incoming radar as different adversaries. Wrote and tested the program in C++ for industrial use after testing concepts in MATLAB.

Identifying airways in lung CT scans and determining their orientation

Occlusion Penetrating Imaging

The goal of this image was to be able to take spy plane images, and track partially occluded targets. Taking multiple images of a scene, camera locations, and object distance, this program will stitch together the scene behind the occluding object (often foliage) by overlaying the images with the pixel displacement to match the distance given and the camera positions. I tested this algorithm by taking images on my own cell phone of occluded scenes, and calibrating them to eliminate rotation or translational inaccuracies. I wrote the calibration and stitching programs in MATLAB.

Source images, after rotation and translational calibration to overcome the jitter that comes with acquiring these images on a cell phone.

This video focuses the images at a specific distance, by moving them on top of each other and averaging the pixel values. Partially occluded objects can be made out because they are usually not occluded in every single image. The "College Park" text and building become much more clear behind the bush. The displayed video is the overlayed images with the depth given at the top of the image in millimeters.

This program in MATLAB that read the data and created scatter plots to model where the feet were on the force plate over time. In this particular trial, the person was instructed to run right. Before running right, the person leans left, which follows the inverse pendulum model.