Face tracking by fusing convolutional neural networks and particle filtering

Objective To address the problem of target loss due to the change of color, shape, texture and other features in face tracking, this paper proposes a face tracking algorithm integrating convolutional neural network and particle filtering. Methods Firstly, a large number of face images with different poses, different lighting and occlusion conditions are trained in the convolutional neural network to get the face feature maps under different conditions, then the trained network weights are used as the weight coefficients in the particle filtering algorithm, and finally the particle filtering algorithm is used for face tracking. Results In this paper, the OPE (one pass evaluation) evaluation criterion in OTB2015 database is used, and the area under the line (AUC) and the center point position error of the success rate graph are used as the evaluation criteria, and by doing comparison experiments with classical KCF, CSK, SRDCF, ECO-HC and other algorithms, the algorithm proposed in this paper achieves The proposed algorithm achieves a tracking success rate of 0.978 and a tracking accuracy rate of 0.97 on the test data. Conclusion In this paper, a face tracking model incorporating convolutional neural network and particle filtering is proposed, which can automatically track regions with large face differentiation and effectively improve the tracking accuracy of the target. Experimental results show that the algorithm in this paper can adapt to face tracking in complex situations such as different postures, different lighting and occlusion, with good robustness and real-time performance.