Scale Invariant Tracking of Objects and Occlusion Detection in Challenging Thermal Infrared Sequences

A major challenge in computer vision field is to track a target in different dynamic conditions such as changes in visual appearance, scale, occlusions and turbulence. Most of the vision-based trackers fails with changes in object appearance and low contrast conditions, especially in thermal infrared (IR) sequences. This paper proposes a robust technique that caters major challenges in tracking of objects in thermal IR sequences and shows promising results on Visual Object Tracking benchmarks (VOT). The proposed technique employs Discriminative Correlation Filtering with spatial and reliability maps to detect tracker failure conditions. It utilizes the Kalman filter for target translation estimation and tracking in situations such as low contrast or object size changes. Hence, using the enhanced correlation based discriminative filter along with occlusion detection and Kalman estimator results in excellent performance on challenging thermal IR sequences. Experiments have been conducted on VOT-TIR-2016 benchmark and CAMEL object tracking dataset to demonstrate the efficiency of the proposed approach in thermal IR sequences. Experimental results proved that our scheme performs better than other approaches by achieving an improved accuracy of up to 0.60, with a change in scale and contrast of the visual objects, whereas the original CSRT had an accuracy of 0.48