| 摘要: | 近年來,深度神經網路獲得了爆炸性的發展與成就,尤其是卷積神經
網路,在影像辨識上已經被成功運用在許多的領域,甚至像是利用三維的
卷積神經網路,利用在人類的手勢辨識[1]、動作辨識[2]等應用。
脈衝神經網路,被稱為是第三代的神經網路,旨在模擬真實的生物神
經細胞,利用二元的脈衝訊號來傳遞訊息,相較於目前一般的神經網路所
使用,經過高度簡化的神經元模型,脈衝神經元更貼近生物神經元的運作
性質。
近年來,動態視覺傳感器的運用正在發展起來。相較於一般基於幀的
相機,DVS 相機是基於事件的相機,只捕捉發生亮度變化的像素點,更接
近人眼感知世界的機制。因此,本論文也利用了 DVS 相機,自行收集了
動作辨識資料集。
本論文基於[3],利用三維卷積脈衝神經網路,分別在 DVS128 資料集
和自行收集之資料集,進行動作辨識的實驗。;In recent years, deep neural network(DNN) has accomplished explosive developments and achievements, especially convolutional neural network(CNN), which has been successfully used in many fields in image recognition. Moreover, we are able to utilize three dimensional CNN to devise some applications, such as human gesture recognition [1], action recognition [2], etc.
Spiking neural network(SNN), known as the third-generation neural network, is designed to simulate real biological nerve cells, which uses binary spikes signals to transmit information. Compared with the highly simplified neuron model used in current general neural network, the spiking neuron model is closer to the operational nature of biological neuron.
In recent years, the use of dynamic vision sensor(DVS) is growing. Compared to general frame-based camera, DVS camera is event-based camera that only captures pixels with brightness changes , which is closer to the human eye′s mechanism of perceiving the world. Therefore, we also used DVS camera to collect our own action recognition dataset for this paper.
Based on [3], we would utilize 3D spiking convolutional neural network(3DS-CNN) in this paper, doing experiments for action recognition on DVS128 dataset and the dataset collected on our own. |
