Accelerating Video Analytic Processing on Edge Intelligence

The most demanding Artificial Video analytic applications require in-edge inference of the AI model to ensure low-latencies to obtain the result. In general, the training process of the model can be executed on the cloud taking benefit from the high-performance computing capabilities available on those premises. This work presents an AI Video analytic application implemented on an Edge-computing device. This device is capable of accelerating the inference of AI models and Video compression by dedicated hardware. This paper presents the architecture designed to implement Image, Networking and Deep Learning Processing functionalities on a reconfigurable System-on-Chip. Additionally, the design tools and design flow followed to generate all software and hardware configuration is detailed. This Edge Intelligence platform is currently in-service, providing the preliminary results for the targeted applications. The proposed solution can process 33 times more video data volume in real-time than the software GPU accelerated implementation for the testing conditions described in the paper.