The SAE AutoDrive challenge is a collegiate design series to transform a GM Bolt EV to a SAE level-4 self driving car. I am honored to be a part of the Michigan Tech team, which is one of eight universities competing in this challenge.

I work on all aspects of the vehicle from sensor interface to software development (perception and sensor fusion) and testing. Year-1 of the competition saw us interface sensors like camera, Lidar and INSS and manage wiring and routing of these sensors. We demonstrated the use of cameras for tasks like lane detection and the use of our INSS system for navigation. For Year-2 of the competition, we demonstrated Lidar point-cloud based object detection and tracking algorithm and a fusion with cameras to achieve pedestrian tracking. We also started to use artificial intelligence for traffic sign and light detection. For Year-3 of the competition, we interfaced automotive radar and are in the process of developing object tracking and track level fusion with other sensors. To compensate for a loss of GPS, we also demonstrated the use of Lidar point-clouds as a source of odometry and achieved multi sensor fusion for simultaneous localization and mapping.

I have been involved in many cross-functional teams and mentored senior year students with their capstone projects over the course of 4 years.

Winter is particularly harsh in the North where phenomenon like arctic polar vortices and lake effect snow bring heavy wet snow often resulting in "white-out conditions" severely reducing visibility and resulting in challenging driving conditions. In this project we attempt to model snow and understand characteristics that will help mitigate negative effects presented by snow. 

Most rotating automotive LiDAR have dual return filters, to select signal returns with the highest remission. This results in a poisson distribution of active snow points with the bulk concentrated close to the sensor (~15m). To effectively filter out such snow, it is important to use an adaptive approach which is aggressive closer to the sensor and relaxes as the distance of a point increases from the sensor.

Details of our work can be found here: https://arxiv.org/abs/2109.07078    & 

Winter adverse driving dataset for autonomy in inclement winter weather

Several publicly available datasets are being actively used to develop autonomous systems that include 2D as well as 3D information. However, very few exist that include data collected in truly inclement weather, across multiple seasons with heavy snow/whiteout conditions. Michigan Tech is situated in the Keweenaw peninsula in Upper Michigan where we receive 200+ in of snow in the winter. This snow often includes polar vortices and lake effect snow from the great lakes that result in challenging driving conditions.

Our aim here is to publish a complete dataset with LiDAR pointclouds, visual information from multiple cameras and INSS information to help promote the development of autonomous systems and algorithms that are able to operate in such conditions.

https://paperswithcode.com/dataset/wads

https://digitalcommons.mtu.edu/wads/

Off road navigation demands ground robots to traverse complex and often changing terrain. Classification and assessment of terrain can improve path planning strategies by reducing travel time and energy consumption. In this project I develop a terrain classification and assessment framework that relies on both exteroceptive and proprioceptive sensor modalities. The robot captures an image of the terrain it is about to traverse and records corresponding vibration data during traversal. These images are manually labelled and used to train a support vector machine (SVM) in an offline training phase. Images have been captured under different lighting conditions and across multiple locations to achieve diversity and robustness to the model. Acceleration data is used to calculate statistical features that capture the roughness of the terrain whereas angular velocities are used to calculate roll and pitch angles experienced by the robot. These features are used to train a k-means clustering classifier, where k is the number of terrain types the robot is expected to encounter. Human intervention is required only to set k. Cluster indices are auto labelled during training. In the operating phase, the SVM predicts the terrain that the robot will encounter and the corresponding cluster center predicts the vibration features associated with that terrain type. These features can be used to assess traversability. Vibration features are measured and the clusters are updated as the robot traverses, thus adapting to change in terrain over time. An accuracy of 92% has been achieved using visual data to classify terrain and 76% to predict the physical properties across different terrain classes. 

Our group undertook the task of upgrading the Front-end Hadron Radiation Monitoring (HFRADMON) system to be installed in the CMS detector at the Centre for European Nuclear Research (CERN). As a part of this project, we developed two modules: an FPGA card (digital sub-system) and an ADC card (analog sub-system).

The FPGA card acquires pulses from the Neutron detectors and activates counters to record the rate of the Neutrons hitting the detectors. The ADC card monitors the voltage and current of the detectors to estimate the health of the detectors. It also monitors the temperature and on-board voltage for housekeeping.

The hardware design of the two cards was extremely challenging due the complexity of the circuits. It was done completely in-house. The FPGA card has been designed on an 8-layer PCB whereas the ADC card has been designed on a 4-layer PCB. The wiring, testing and debugging of the complete system was done by me. It has been an interesting task and extremely educational.

All the communication on the ADC Board is done using the I2C protocol. The ‘rate’ data (from counters) along with the health information and housekeeping data is sent to the PC by the FPGA card over an Ethernet link, using the UDP protocol.

I also developed a Backplane PCB, similar to the CERN setup, so as to completely emulate the conditions. This has greatly improved the connectivity issue we faced during prototype. We were also able to test all the possible conditions that could arise during final operation. The below figure shows the complete setup (FPGA card, ADC card and Backplane Connector card) housed in an enclosure.

Silicon Photo-Multipliers (SiPM) are superior photo detectors as compared to conventional sensors. Immunity to magnetic field, compact footprint, high photon detection efficiency (PDE) and low bias voltage are some key features of an SiPM. However, they are highly temperature sensitive. The gain of an SiPM varies by almost 3-5% / ˚C. To effectively use this device in high energy physics experiments, like GRAPES-3, where the ambient temperature varies by almost 10-15 ˚C, the bias voltage must be adjusted dynamically with respect to the temperature.

Our group has developed a temperature compensated Programmable Power Supply (PPS) to stabilize the gain of SiPM. This PPS has the capability to bias 16 SiPM's at once and can also measure the current through them.

User can set an appropriate voltage by communicating via. a USB-2.0 link. A  micro-controller (PIC18F family) is the main processing unit of the system. A 14-bit DAC is used to provide reference voltage to a series of linear regulators to generate the required voltage. The PPS automatically adjusts the bias voltage with a temperature feedback mechanism with a TSiC-501F temperature sensor that provides an accuracy of 0.1 ˚C.

I undertook the calibration and validation of the PPS system. The various components used on the board show slight non-linearities making the calibration of the system extremely important. After the calibration process the system was validated by putting it through rigorous testing for accuracy, resolution, stability and driving capacity.

Details of calibration and validation and results have been discussed in detail in our paper which has been submitted to the Review of Scientific Instruments.

Over a temperature sweep of 15 ˚C, the SiPM gain was observed to fall by ~50%. However, with our temperature compensation, the SiPM was observed to maintain the gain within 0.5% over the complete range of 15 ˚C. The below figure shows the normailzed gain of an SiPM with and without compensation.

The output signal of the SiPM is of the order of a few hundred uV. Therefore, in order to apply any further processing, it is necessary to boost the voltage level. Hence we undertook the design of a high-speed low-noise amplifier. We explored both fixed gain as well as variable gain amplifiers in single stage as well as multi stage configurations.

We were able to design an amplifier card in an extremely small factor (1.5cm x 2.5cm). The histogram in the below figure shows the response of the amplifier card at high laser intensity (large no. of photons incident). The peaks correspond to different photo-electrons. The spacing between these peaks indicate the high gain of the amplifier and the good fit shows the low noise feature of the circuit.