A Study on the Impact of Multiview Distributed Feature Coding on a Multicamera Vehicle Tracking System at Roundabouts

Abstract

Visual sensor networks are one potential enabler for the evolution of the Internet of things. Due to their limited resources in terms of energy and bandwidth, it is crucial to identify appropriate approaches that take into considerations such constraints and reduce the amount of data transmitted to the gathering point (sink). In this context, this paper describes the impact of a distributed smart-camera system that exploits an analyze-then-compress strategy, on a multi-view vehicle tracking at roundabouts application. In the tested system, part of the processing is shifted to the smart cameras, i.e., the object detection/classification and feature extraction, so that only the extracted features describing moving vehicles are transmitted instead of the whole image/video. Features are further compacted by using a state-of-the-art distributed coding technique, based upon an efficient clustering method that exploits the temporal and spatial (multiple views) correlations between features. The system is tested on a real-data scenario, by evaluating the bit-rate reduction capabilities in dependence of the channel conditions, as well as the matching accuracy of the reconstructed descriptors in the specific tracking application. Both feature-wise and object-wise matching are investigated. For the chosen application scenario, a bit-rate reduction of 30 – 35% is proved to be achievable in non-ideal channel conditions. Even more interestingly, such reduction is proved not to harm the matching accuracy (i.e., it is coherent with the target application), for which an F-score up to 0.923 is guaranteed.