IMPROVEMENT OF PERFORMANCE OF HUMAN DETECTION IN ABNORMAL CROWD USING KNOWLEDGE DISTILLATION FOR YOLO NETWORKS

Abstract

Deep Neural Networks has been achieved outstanding advantages in many different fields, especially object detection using the YOLO networks. The YOLO models
have increasingly improved to obtain efficiency and overcome shortcomings of previous versions. However, in order to obtain better performance that the network
structure is more complex, the number of model parameters is larger, and requires the longer response time and inverse. Especially, human detection problem in
abnormal crowds that faces to some problems such as the high density of people in a frame and the lager speed of human movements. To achieve high accuracy in
human detection of anormaly crowd context, the model requires a complex architecture, which leads to high time cost. In this study, the knowledge distillation from
multiple higher configurations of the complex YOLO models to a simpler configuration of YOLO model. Experimental results performed on two unusual crowd databases
show that our propose framework achieves better results on both single database evaluation and cross database evaluation from 1% to 6.8% higher in accuracy, the
response time of our method reduces 9.16ms than YOLO V8 when it tested on GPU.