Abstract: In order to deal with scale variation, illumination changes and deformation of the target in tracking tasks, a visual tracking algorithm based on mixed features, called SoH-DLT, was proposed, considering both the contour features and detail features. Orientation histogram is introduced to describe the contour features of candidate samples in the process of particle filter, ensuring that the particle which is the most similar with the target can still get a high degree of confidence and can be output as the result of tracking in the case of scale, illumination changes and deformation. Speed-up robust features (SURF) feature and high-level features from deep learning are integrated to calculate the weights of particles, improving the tracking accuracy in complex scenes and enhancing the robustness of SoH-DLT to scale variation. Experiments show that SoH-DLT algorithm has better tracking performance than the contrast algorithms in both quantitative and qualitative evaluation.

Key words: visual tracking, contour feature, neural network, orientation histogram, speed-up robust features, particle filter