Approximation for streaming video analytics on mobile devices

Approximate algorithms have shown great success to reduce the computation latency with minor accuracy loss. Such algorithms are especially useful on resource-constrained devices like mobile phones and embedded boards. However, current approximate algorithms for streaming video analytics cannot provide accuracy and latency guarantees. This happens in the face of changing content characteristics in the video stream and changing resource availability on the target devices.

In this dissertation, we propose the analytic foundation on how to characterize the accuracy and latency of approximate algorithms in a content and contention aware manner. We use this to enable an approximate algorithm to probabilistically meet a user-provided latency constraint or an accuracy target. We use video processing pipelines, a video object classification system, and a video object detection system as examples to demonstrate how our solution can improve the performance of streaming video analytic systems on the resource-constrained mobile and embedded devices. Our evaluation shows that our technique can be overlaid seamlessly on top of standard vision algorithms and provides superior accuracy-latency tradeoff over the start-of-the-art approaches.