Reducto: On-Camera Frame Filtering for Resource-Efficient Real-Time Video Analytics
Real-time queries of live video can guide numerous applications to find useful information at a timely fashion. To extract such information from videos, analytics pipelines rely on deep neural networks for tasks such as object detection. Unfortunately, such DNNs impose high resource demands on backend servers, which often prevent analytics pipelines from delivering low-latency responses for real-time queries. Recent techniques reduce backend computation by employing edge servers to filter out irrelevant frames using lightweight NN models. However, edge servers pose numerous challenges with respect to deployment costs and privacy, and they ignore potential network bottlenecks at cameras.
This paper presents Reducto, a camera-server system that provides low enough latency for answering real-time video queries. The key idea behind Reducto is filtering out unnecessary frames directly at the data source (i.e., cameras) to achieve additional efficiency. However, commodity cameras often lack the resources (e.g., GPUs) required to run even lightweight NN models. Our key observation is that techniques based on differencing features are very cheap to run and can effectively track frame differences, which are well correlated with changes in query results. Based on this observation, we built Reducto that can effectively filter out frames at the camera based on the best feature selected by the server for each query. Due to use of cheap techniques, the camera can constantly check the validity of the feature and once it grows out of date the server quickly selects a new one. As such, Reducto ensures that the filtering technique adapts to the inherent dynamism of the video and meets the desired accuracy at a fine granularity. Comparisons between Reducto and existing systems demonstrate very promising results.