Real-Time Interpretation Method for Shooting-Range Image Based on Position Prediction

Abstract

When the theodolite in the shooting-range tracks the target in real time, the theodolite may randomly jitter, causing the target a large displacement on the next frame. While dealing with a large displacement, the tracking methods based on the window search are easy to lose the target, and the tracking methods based on the full-image search are time-consuming. In this paper, an improved tracking-learning-detection (TLD) framework combining Kernelized Correlation Filters (KCF) and target position prediction is proposed to cope with the large displacement. First an orthogonal polynomials optimal linear filter is used to predict the position of the target on the next frame according to the rule of the theodolite’s angle, and then the KCF is used to track the target fast in this prediction area, which can improve the success rate and speed of tracking. If the tracking fails and the prediction area is in the image, the detector will detect the target in the full image. Simulation experiments have demonstrated that the position prediction algorithm based on the optimal linear filter can accurately predict the target position and provide KCF with a more accurate search position. The algorithm consumes only 0.6 ms per frame, and the tracking accuracy is better than TLD and KCF. The actual task verification of the shooting-range has proved that our method can improve the automatic interpretation of the shooting-range and reduce manual intervention.