基于人类记忆机制的码本建模方法研究

码本模型在实际应用时需要提前进行离线训练构建码本,在一定程度上限制了码本模型的实时应用。为此,将人类三阶段记忆机制引入到码本建模过程,基于记忆机制进行在线码本建模。每个像素点的时序信息被抽象成码本中不同的码字,并通过在瞬时、短时和长时记忆空间模拟记忆、遗忘、回忆等认知行为,实时地对码字进行分类。实验结果表明,提出的新算法不仅提高了模型的实时性而且对场景突变的适应能力也强于原始模型;同时,新算法在交通监测等现实应用中也能取得良好的效果。

Memory-based codebook modeling for real-time object detection

Codebook modeling builds codebook through offline training before segmentation, which will affect its real-time application to some extent. In order to tackle this problem, in this paper, three-stage memory mechanism is introduced into the construction and updating process of codewords and a memory-based codebook model (MCB) which is capable of online learning is proposed. The sequential information of each pixel is abstracted into codewords which are classified by imitating some cognitive behaviors such as memorization, forgetting, and recall. Experimental results demonstrate that our method not only improves real timing but also is much more adaptive to the scenes with sudden changes than the original model. In addition, the proposed method can be applied in traffic monitoring.