CCT: Context and Colorization for Tracking

Video-based multi-camera tracking has witnessed tremendous progress in recent years. Deep learning has led to dramatic gains in speed and accuracy of tracking algorithms. Similar success is seen in RF-based and other single sensor-based trackers. Multi-sensor tracking however, remains a challenge. The problem is especially difficult when sensors provide detections of unknown confidence, which makes it difficult to leverage the successes of existing Bayesian approaches to tracking. To address this problem, we propose a deep learning-based semi-supervised approach which learns to map heterogeneous features in a common embedding space. Comparative evaluation will be used to demonstrate the benefits of the approach.