Oct 31, 2017 (modified: Oct 31, 2017)NIPS 2017 Workshop MLITS Submissionreaders: everyone
Abstract:Video image processing of traffic camera feeds is useful for counting and classifying vehicles, estimating queue length, traffic speed and also for tracking individual vehicles. Even after over three decades of research, challenges remain. Vehicle detection is especially challenging when vehicles are occluded which is common in heterogeneous traffic. Recently Deep Learning has shown remarkable promise in solving many computer vision tasks such as object recognition, detection, and tracking. We explore the promise of deep learning for vehicle detection and classification. However, training deep learning architectures require huge labeled datasets which are time-consuming and expensive to acquire. We circumvent this problem by data augmentation. In particular, we show that by properly augmenting an existing large general (non-traffic) dataset with a small low-resolution heterogeneous traffic dataset (that we collected) we can obtain state-of-the-art vehicle detection performance. This result is expected to further encourage the wide-spread use of deep learning for traffic video image processing.
TL;DR:We have provided a new dataset for vehicle detection in heterogeneous traffic and have also proposed a method of data augmentation in deep architectures for the same.