Handling Adverse Visual Conditions for Target Tracking and Recognition

Handling Adverse Visual Conditions for Target Tracking and Recognition Extending visual sensing capacity in target tracking and identification is very important for Army’s future force. Contemporary techniques assume and depend on good quality imagery. In practice, however, as the environments are unconstrained, their performances are largely degraded in adverse visual conditions, e.g., induced by bad weather and illumination conditions, when the visual quality of the data is seriously degraded. We call them perceptually ”low-quality” imagery. The existing solution to this issue is to perform pre-processing to restore the quality of the imagery, e.g., via image super-resolution or de-blurring. However, such image restoration tasks themselves are very difficult and computationally demanding. Therefore, this solution is not practical. The goal of this project is to explore an innovative solution to overcome this challenge, by exploring a unified approach that does not perform explicit image restoration as pre-processing in target tracking and recognition under various adverse visual conditions and degradations. Specifically, we plan to address the following issues: � A principled approach and its theoratical foundation. The key idea of avoiding performing explicit image restoration is to embed the prior knowledge for restoration into visual target matching. Our objective is to develop a general approach of learning image similarity and visual regression that applies to various situations of adverse visual conditions. � Target matching and tracking. The performance of target tracking is largely determined by the visual similarity metric. It should be adaptive to different imagery. Our objective is to learn the metric for the low-quality imagery under adverse visual conditions, by steering and aligning the known metric from good-quality images. � Target identification. Targets are seen from very different views. It is very difficult to determine if the low-quality images from different views are the same target. Our objective is to learn the visual similarity metrics to generate the predicted novel views. The difference between the actual image and predicted view will be used for better target identification. � Target attribute estimation and recognition. A target has various descriptive attributes. Some are discrete and others are continuous. It is very difficult to extract them from low-quality images. Our objective is to estimate and recognize those attributes from low-quality imagery via learning reconstruction-based visual regression. One innovation of the proposed approach is that it avoids performing explicit, expensive and dedicated image/video restoration which is in general computationally demanding, but rather using the low-quality data directly with the implicit knowledge learned from data. This research has not been studied before, and it leads to innovative and computationally efficient solutions to track and recognize targets in adverse visual conditions. Moreover, it is a principled and general solution that will be able to handle various image degradations in the same framework. Once we have collected the training data (i.e., an exemplar database) that associate the low-quality data with their corresponding high-quality data, it conveys the implicit knowledge that facilitates to learn the similarity metric of the low-quality data, and it will empower effective reconstruction-based local regression. This will make possible target matching, tracking and recognition on low-quality data