Federated Video Analytics System Design

[Sep. 2018 - Apr. 2020]

• Compared different object detectors for embedded devices and studied the trade-off between inference accuracy and latency with addressing the optimization of both computing and networking in the edge computing system
• Designed algorithm for optimizing video analytics performance (accuracy/end-to-end latency of object detection) by leveraging on-device processing and computation offloading in edge computing system
• Applied black-box optimizer to select the object detection neural network model running on embedded devices, frame resize rate, and the number of frames for offloading
• Trained neural process to learn the distribution of system performance based on network status and system configurations
• Built testbed with using NVIDIA Jetson TX2, nvidia-docker(GPU-enabled docker), OpenWRT router, and desktop GPUs
• Implemented algorithms in Python to test the proposed video analytics pipeline and evaluate end-to-end performance of the system