Consistency in Multi-modal Automated Target Detection using Temporally Filtered Reporting

摘要

Autonomous target detection is an important goal in the wide-scale deployment of unattended sensor networks. Current approaches are often sample-centric with an emphasis on achieving maximal detection on any given isolated target signature received. This can often lead to both high false alarm rates and the frequent re-reporting of detected targets, given the required trade-off between detection sensitivity and false positive target detection. Here, by assuming that the number of samples on a true target will both be high and temporally consistent we can treat our given detection approach as a ensemble classifier distributed over time with classification from each sample, at each time-step, contributing to an overall detection threshold. Following this approach, we develop a mechanism whereby the temporal consistency of a given target must be statistically strong, over a given temporal window, for an onward detection to be reported.rnIf the sensor sample frequency and throughput is high, relative to target motion through the field of view (e.g. 25fps camera) then we can validly set such a temporal window to a value above the occurrence level of spurious false positive detections. This approach is illustrated using the example of automated real-time vehicle and people detection, in multi-modal visible (EO) and thermal (IR) imagery, deployed on an unattended dual-sensor pod. A sensitive target detection approach, based on a codebook mapping of visual features, classifies target regions initially extracted from the scene using an adaptive background model. The use of temporal filtering provides a consistent, fused onward information feed of targets detected from either or both sensors whilst minimizing the onward transmission of false positive detections and facilitating the use of an otherwise sensitive detection approaches within the robust target reporting context of a deployed sensor network.