Pedestrian detection system for smart communities using deep Convolutional Neural Networks

Jonathan Lwowski, Prasanna Kolar, Patrick Benavidez, Paul Rad, John J. Prevost, Mo Jamshidi

2017 · 29 citations · 15 references

DOIFull text

Open access

Concepts

TL;DR

Pedestrian recognition is essential for autonomous driving, search and rescue, surveillance, and robotics, yet real‑time detection is hampered by the high computational cost of region‑based CNNs such as Fast R‑CNN, which, while excellent at image classification, suffer from slow inference. This work proposes a fast, accurate pedestrian detection system that cascades a lightweight regional detector with a deep convolutional network to enable real‑time alerts for drivers. The approach combines a rapid regional detection stage with a deep CNN, achieving real‑time performance on a low‑performance system without a GPU. The system attains 95.7 % accuracy while processing approximately 15 frames per second on a non‑GPU platform.

Abstract

Pedestrian recognition is a key problem for a number of application domains namely autonomous driving, search and rescue, surveillance and robotics. Real-time pedestrian recognition entails determining if a pedestrian is in an image frame. State-of-art pedestrian detection convolution neural networks(CNN) such as Fast R-CNN depend on computationally expensive region detection algorithms to hypothesize pedestrian locations. This paper presents a simple, fast and very accurate approach by cascading fast regional detection and deep convolution networks. Convolution networks have been shown to excel at image classification. However, convolution networks are notoriously slow at inference time. In this work, we introduce a fast regional detection cascaded with deep convolution networks that enables real-time pedestrian detection that could be used to alert a driver if a pedestrian is on the roadway. The classification CNN has given an accuracy of 95.7%, with a processing rate of about 15 frames per second on a low performance system without a graphical processing unit (GPU).

References

15