Benchmarking General-Purpose Neural Networks for Real-Time Pedestrian Detection

Abstract

This paper presents the main steps and results on automotive applications aiming to detect pedestrians by means of object detection and recognition neural networks. The used training and testing datasets take into account the influence of different factors, such as environmental conditions and video-image acquisition characteristics, using a custom training and testing dataset, and critically evaluate the performance and capabilities of general-purpose neural networks in detecting pedestrians from usual images provided by a video camera.Following this goal, an analysis points out what differences occur in the final classification output of these neuronal networks using the same dataset for training while testing on real-life scenarios. In the end a general-purpose neural network which provides the best results in pedestrian detection (providing 2D bounding boxes) is proposed and its performances are discussed.