HPDC group, Vrije Universiteit Amsterdam
EDL P16-25 P1: Deep Learning as a Service (DLaaS)
Services for Large-scale On-line Surveillance
My research focus is on designing and implementing efficient systems/services for real-time video stream analytics. Video stream analytics has become a core component of many applications such as augmented reality, public safety., traffic monitoring, etc. We now do get somewhat higher accuracy from DNN-based computer vision algorithms compared to the traditional algorithms. But the increased accuracy comes at the cost of the increased computational complexity of DNNs due to deep and complex architechtures, which might hinder the real-time objective, for example, processing 30FPS. Therefore, the question that we try to answer is, ‘how to design a system that supports efficient online video analytics?’
More specifically, in our first paper, we have worked on the question, ‘how and where to deploy deep learning models to identify human actions in real-time on live video streams?’. Unlike many existing approaches that propose to use edge-only (i.e., a resource-constrained device available near the camera source) or cloud-only system (i.e., a powerful remote server), we proposed a hybrid system that leverages the benefits (fast response and high accuracy) of both the system by deploying a smaller model on the edge and a bigger model in the cloud.
Currently, we are working on designing a system that guarantees SLA (e.g., satisfies latency constraints without affecting the analytics accuracy) when DNN models are accessed over a variable communication network. Hint: By adapting DNN models during runtime. 😉
Clownfish: Edge and Cloud Symbiosis for Video Stream Analytics. Vinod Nigade, Lin Wang, Henri Bal. ACM/IEEE Symposium on Edge Computing (SEC), 2020. Conference proceedings, Open access-green.
Better Never Than Late: Timely Edge Video Analytics Over the Air. Vinod Nigade, Ramon Winder, Henri Bal, Lin Wang. ACM SenSys AIChallengeloT. Conference proceedings, Open Access-Gold.
| Github |