Visual Tracking Based on Convolutional Deep Belief Network

Visual tracking is an important task within the field of computer vision. Recently, deep neural networks have gained significant attention thanks to their success on learning image features. But the existing deep neural networks applied in visual tracking are full-connected complicated architectures with large amount of redundant parameters that would be low efficiently to learn. We tackle this problem by using a novel convolutional deep belief network (CDBN) with convolution, weights sharing and pooling to have much fewer parameters to learn, in addition to gain translation invariance which would benefit the tracker performance. Theoretical analysis and experimental evaluations on an open tracker benchmark demonstrate our CDBN based tracker is more accurate by improving tracking success rate 22.6% and tracking precision 62.8% on average, while maintaining low computation cost by reduces the number of parameters to 44.4%, compared to DLT, another well-known deep learning tracker. Meanwhile, our tracker can achieve real-time performance by a graphics processing unit (GPU) speedup of 2.61 times on average and up to 3.08 times.