Research on target tracking algorithm based on multi-layer convolution feature fusion
Kehao Xiao, Hao Zhou · 2022
Due to the poor tracking performance brought on by external factors like the target objects moving quickly, changing light, and scale fluctuation, the Convolutional Neural Network (CNN) trained on the target identification dataset was utilized for feature extraction. On the foundation of this, a multi-layer convolution feature fusion-based target tracking method was developed. First, the VGG19 was used to execute the multi-layer convolutional feature extraction. Second, to finish the target position estimate based on the multilayer convolutional features, the correlation filter tracker was built individually for each layer of convolutional features, and the features were fused by assigning weights according to the feature responses to complete the target position estimation based on the multilayer convolutional features. Once multi-scale sampling had been performed in the target center region in accordance with the correlation filtering model, the scale variation of adjacent frames was ultimately employed to complete the scale prediction of the target. The tracking performance was empirically examined in the dataset OTB50 and contrasted in numerous ways using various tracking methods. The approach presented in this work significantly improved tracking rate and accuracy while exhibited good robustness to scale variation.