Object Tracking and Classification in Videos Using Compressive Measurements

In this paper, we summarize some recent results on objective tracking and classification in infrared and low quality videos using compressive measurements. Two compressive measurement modes were investigated. One was based on subsampling of the original measurements. The other was based on coded aperture camera. It is important to emphasize that conventional trackers require the compressive measurements be reconstructed first before any tracking and classification processing steps begin. The reconstruction is time-consuming and may also lose information. Our proposed approach directly uses compressive measurements and a deep learning tracker known as You Only Look Once (YOLO), which is fast and can track multiple objects simultaneously, was used to track objects. The detected objects are then recognized using another deep learning model called residual network (ResNet). Extensive experiments using infrared videos from long distances were conducted. Results show that the proposed approach performs much better than conventional trackers, which failed to deal with compressive measurements. Instead, ResNet classifier performs better than the built-in classifier in YOLO.