Efficient Deep Network Architecture for Vision-Based Vehicle Detection Keyvan Kasiri,

Abstract

With the progress in intelligent transportation systems in smart
cities, vision-based vehicle detection is becoming an important issue
in the vision-based surveillance systems. With the advent of
the big data era, deep learning methods have been increasingly
employed in the detection, classification, and recognition applications
due to their performance accuracy, however, there are still
major concerns regarding deployment of such methods in embedded
applications. This paper offers an efficient process leveraging
the idea of evolutionary deep intelligence on a state-of-the-art deep
neural network. Using this approach, the deep neural network is
evolved towards a highly sparse set of synaptic weights and clusters.
Experimental results for the task of vehicle detection demonstrate
that the evolved deep neural network can achieve a substantial
improvement in architecture efficiency adapting for GPUaccelerated
applications without significant sacrifices in detection
accuracy. The architectural efficiency of ~4X-fold and ~2X-fold
decrease is obtained in synaptic weights and clusters, respectively,
while the accuracy of 92.8% (drop of less than 4% compared to the
original network model) is achieved. Detection results and network
efficiency for the vehicular application are promising, and opens
the door to a wider range of applications in deep learning.