Implement tracking algorithm using CNNs

Convolutional neural networks (CNNs) is widely used as classifiers in the field of computer vision. The more complex a CNNs model is, the more accurate classification results will be. But a very deep network also requires a better GPU to train and test in a reasonable time. In this paper, we purpose a tracking algorithm using LeNet-5, and avoid computing complex handcrafted features from the raw inputs. Modifying the number of kernels improves the tracking accuracy of models, avoid over-fitting. The experiment of processing a video clip meets real-time requirement while using GPU and it shows our algorithm is more robust than traditional algorithm like particle filter to track single target under the complicated background.