The Research Of Outdoor Vehicle Visual Tracking Algorithm Based On Deep Learning

Video object tracking is an important research in the field of computer vision, which is widely used in intelligent transportation system. And the outdoor vehicle tracking is an important component in it. However, due to the complex and changeable outdoor environment,such as illumination variation, occlusion, and background clutters, the outdoor vehicle tracking is facing severe challenges. So this paper proposed 3 deep-auto-encoder-based algorithms under particle filter framework for outdoor vehicle tracking, and put convolutional neural network into outdoor vehicl tracking.(1) The k-sparse-deep-denoising-auto-encoder-based object tracking algorithm. This algorithm extracted object feature that can effectively distinguish object from background by k sparse denoising auto-encoder. The whole algorithm was consisted of offline training,classification neural network construction and online tracking. Firstly, this algorithm offline learned generic feature representation by training samples, and then constructed classification neural network by the learned knowledge. Finally the output confidence of classification neural network was used in online tracking.(2) The multi-deep-auto-encoder-fusion-based object tracking algorithm. Pointing at the situation that gray-scale feature is easily affected by illumination changes in algorithm (1), this algorithm put multi-deep auto-encoder into object tracking. The whole algorithm was consisted of generic feature representation, classification neural network fusion and online tracking.Firstly, this algorithm employed gray-scale feature and gradient feature of 300 000 images which is selected from tiny images dataset randomly for two deep auto-encoder's offline training, and then constructed and linear weighted fusion two classification neural networks according to the offline training results. Finally, the fusion result was used in online tracking.(3) The multi-deep-auto-encoder-adaptive-fusion-based object tracking algorithm. To be directed against the unreliable multi-model fusion result when the fusion weight is fixed in algorithm (2), this algorithm proposed an adaptive fusion strategy. In the strategy, the adaptive fusion weight was automatic determined flowing by the distribution condition of particles that represented by different models.(4) The convolutional-neural-network-based object tracking algorithm. This algorithm was consisted of offline training and online tracking. Firstly, this algorithm employed 60 000 images from CIFAR-10 for supervised offline training of convolutional neural network, and then put the training result into online tracking by transfer learning way. The classification neural network of the above four algorithms must be fine-tuned in order to adapt object changes in the first frame and when the confidence can not satisfy the pre-defined threshold in tracking.This paper designed 4 experiments to evaluate the above 4 tracking algorithms.Experiment (1) and experiment (2) evaluated the algorithm performance by quantitative and qualitative way. The quantitative evaluation was conducted by precision and success rate on 50 fully annotated video sequences with overall performance and 11 attributes-based performances in VTB dataset. And the qualitative evaluation was conducted on 4 and 12 challenging outdoor vehicle sequences respectively. Compared with 3 state-of-the-art tracking algorithms (DLT,MTT and CSK), experiment (1) shows that the proposed k-sparse-deep-denoising-auto-encoder object tracking algorithm has higher tracking accuracy in various challenging environmental factors, such as illumination variation, occlusion, background clutters, scale variation, etc.Compared with 7 state-of-the-art tracking algorithms (DLT, IVT, LI APG, MIL, OAB, MTT and CSK), experiment (2) shows that the proposed multi-deep-auto-encoder-fusion-based object tracking algorithm can achieve robust tracking of outdoor vehicle in complex scenes.Experiment (3) compared algorithm (2) and algorithm (3) with center position error and overlap on 6 challenging outdoor vehicle sequences. Result shows that the proposed multi-deep-auto-encoder-adaptive-fusion-based object tracking algorithm can deal with the outdoor vehicle tracking problem that in complex environment better. Experiment (4) was carried out on 3 challenging outdoor vehicle sequences for algorithm (4) evaluation. Result shows that the convolutional-neural-network-based object tracking algorithm does not track the object accurately in some challenging situations. This paper analyzed the reasons furthermore.