An Efficient Pipeline for Distant Person Detection and Identification in 4K Video using GPUs

The gradual advent of machine learning has been assisting to shift the field of computer vision from statistical methods to deep neural networks. These networks should be able to process high resolution video streams coming from the HD camera sources in real time. However, due to the fixed network size and to maintain the processing speed, high resolution frames need to be resized and down-sampled before feeding into the networks resulting in loss of feature information, hampering recognition accuracy. This motivated us to propose a methodology which focuses on creating and processing the active region of interests in the foreground image through an active region generator (ARG) module, eliminating the need to traverse the entire frame and down-sample the resolution before feeding it to the neural network. This resulted in saving 25x more image feature information, whilst maintaining a given person detection accuracy of 92 % mAP for longer distance up to 30~35 metre executing in real time w.r.t it's classical counterpart based on singleshot detector model. Besides, our proposed pipeline architecture utilizing multi-core TESLA GPU increases the execution throughput by a factor of 3X verified in NVIDIA DGX system.