Hybrid Tracking of Low SNR Targets

In general, target tracking follows one of two basic approaches: either the individual scan/stares of an area are thresholded and the detections are tracked, or all the energy received in the scan/stare can be utilized by a track-before-detect (TBD) algorithm. TBD is the preferred approach for tracking targets with low signal-to-noise-ratios (SNR), as thresholding above a set SNR may not detect the target and thresholding below a set SNR may produce far too many detections. There are several different types of TBD algorithms, one of which is the multitarget histogram-probabilistic multiple hypothesis tracker (H-PMHT). However, the issue with methods such as these is that they are primarily track maintenance algorithms and do not have a natural track initiation phase. A hybrid approach seeks to initialize the tracks using low thresholds and then maintain the track with TBD algorithms. It has been found that track initiation (as opposed to track termination) is responsible for discarding tracks unless the algorithm is sure they are valid. This implies that a great deal of false tracks are formed in initiation, and may use extensive resources to monitor and promote to confirmed tracks. Additionally, there are also TBD algorithms such as the Hough Transform, that have the initiation and tracking combined, which are also very computationally expensive. Therefore, to maintain computational efficiency and more accurately initiate tracks, we will introduce smart detections. In this paper we will utilize convolutional neural networks (CNN`s) that examine a chip of pixels surrounding low threshold detections to determine if the energy in the chip came from a target or background noise/phenomenology. The tracks will be initiated with these smart detections and maintained either with future smart detections or the H-PMHT TBD algorithm when the CNN cannot determine that the target is present. The CNN will err on the side of discarding ambiguous detections (meaning real detections will be missed) so that few false tracks are formed. The hybrid approach of using smart detections and PMHT for track maintenance will ensure the target is kept in track even though the CNN may miss detections.