Shrey Mohan Email & Phone Number

Uttarakhand, India

Milpitas, California, Us

Developed activity recognition models to detect actions like entering an O.T room, hand sanitization and exiting an O.T room. The correct sequence of predicted actions helped us to check if the 5 movements of hand-hygiene were being followed by the nurses/doctors or not. Was able to achieve mAP of 92.3% on the in-house validation set.Used yolov5 to predict simple actions like sitting up, lying down and falling. The system was deployed in an elderly care home and achieved an accuracy of 89.3%

Bengaluru , Karnataka , In

Vehicle tracking from drone1. Analyzed the requirements of clients for their task of vehicle testing and formulated constraints for its inspection using a drone.2. Used AWS sagemaker for training object detection model on the UAV (Unmanned Aerial Vehicle) dataset for the task of tracking vehicles being tested in the warehouse.3. Validated and inferenced the trained model using a docker container for deployment on AWS lambda.3D Point Cloud1. Analyzed the bag files recorded using intel real-sense camera, using the PCL library and r-viz application.2. Applied planar segmentation on the point cloud data from the bag file to extract paths to be parsed by the robo-dog.

Charlotte, Nc, Us

Vehicle detection and tracking1. Formulated the FPS and distance constraints for detection and tracking of highway vehicles for developing real-time safety system for short-term highway work-zones.2. Trained and validated Google's efficientdet model for vehicle detection on the UA-Detrac dataset while deploying it on embedded devices.Person Re-identification (Re-id) for scalable tracking on the edge1. Extracted feature maps for individual objects in a video scene using Yolo v3 and used structural similarity index to re-id them in subsequent frames at run-time.2. Trained a triplet loss network with different base networks on the Duke-MTMC, Market-1501 and CUHK03 datasets for re-id with mixed precision for running it on the edge node (Nvidia jetson Xavier). Scalable pedestrian trajectory analysis using LSTM1. Trained a LSTM model using key-points obtained from the OpenPose framework on the Duke-MTMC dataset to predict accurate position of pedestrians in the future frames (next 5).2. The model was trained on single objects (Derived from multi-object videos) and scaled for multiple objects at run-time. The model size was under 1 MB and proved very effective for mobile devices. Real-time person detection and re-id in a camera network1. Collated OpenPose and a triplet-loss (with MobileNet) network for end to end detection and re-id respectively over a network of two edge nodes (Nvidia Xavier) and an edge server (Laptop).2. We kept track of local, global and disappearing objects with the help of local history table at the edge for local re-id and global history tables at the server for global re-id with a stipulated life-span for detections. To protect personal data of people only the extracted features were sent to the server for global re-id.3. To optimize performance the re-id network was ran on the Deep Learning Accelerator (DLA) module of the jetson xavier using the TensorRT framework by Nvidia so that the system could achieve 8 FPS processing speed.

San Jose, California, Us

Using Long Short Term Memory(LSTM) neural network to predict the power of hot boxes, which the company manufactures, for the next quarter after analyzing time-series data(daily granularity) which includes many features like temperature, fuel utilization, average current and voltage consumed, etc.Learning:1. Python(Pandas, SciPy, SciKit-learn), jupyter.2. Deep Learning(LSTM, RNN).3. Keras(Theano, Tensorflow).

Master'S Degree, Computer Engineering

Bachelor Of Engineering - Be, Electrical, Electronic And Communications Engineering Technology/Technician

Doctor Of Philosophy - Phd, Computer Engineering