Target tracking based on improved STRCF algorithm

Target tracking gets great attention in recent years. The correlation filter uses Fast Fourier Transform (FFT) to convert the convolution in time domain to the multiplication operation in frequency domain, thereby effectively training the filter model. The initial tracking frequency based on the Discriminant Correlation Filter (DCF) can reach 700 frames per second. DCF has progressed rapidly in recent years. Trackers such as Spatially Regularized DCF (SRDCF) and Continuous Convolution Operator Tracker(C-COT) have a high degree of accuracy when tracking targets. However, while pursuing better tracking performance, the high-speed and real-time characteristics of the relevant filters are also gradually declined. The increase in the complexity of the model and the variety of target features increases the risk of over-fitting of these trackers. To solve these problems, this paper proposes three solutions: 1. Use deconvolution algorithm to reduce the dimensionality of input image features, thereby reducing the amount of model update operations, improve the speed of our tracker; 2. Prediction of the target position, which reduces the number of candidate boxes, speeds up the positioning process, and improves the tracking performance of moving targets. 3. Reduces the frequency of model updates, saves tracking time, and avoids model drift. Compared with STRCF, our tracker with deep features provides a 5×speedup with only 3.1% decrease in success plots rate (SR) on OTB-2015.