嬰幼童危險預測系統

Abstract

嬰幼兒的安全照護需要大量且密集的人力介入，對於照顧者來說不但費力而 且勞心。大部分的嬰幼兒意外往往是由於照顧者的經驗缺乏或疏失而造成無法彌 補的傷害，此外，若是照護員不當管教甚至虐待嬰幼兒，即使已經加裝監視系統， 也只能事後透過人工方式調閱還原真相，並無法降低傷害或減輕父母的擔憂。為了 能更即時反應嬰幼兒的活動安全狀況，本論文利用電腦視覺技術與結合深度學習， 分析嬰幼兒的行為、所在場域中的人物與物件, 對嬰幼兒的活動狀態進行危險值的 估計，以即時預測並預防意外事件的發生。我們提出了基於遞迴神經網路的嬰幼兒 意外預測模型，稱之為 BabyNet。首先每一幀會進入三個特徵提取模組，取得對應 的特徵後，接著透過遞迴神經網路預測出潛在的危險值。為了解決不同樣本中可預 測性的不同，本文所提出的自適性損失函數根據影片的光流估計出影片的運動能 量，以此作為調整權重的依據。為了實驗及訓練，我們搜集並建立了嬰兒意外的資 料集，總共包含 1200 部的影片。根據實驗結果，BabyNet 預測的平均準確率高達 88.6%，而在 AUC 的表現也高達 0.92，除此之外還能在夠在 3.671 秒前成功預測 意外發生。

Child injuries are a large and growing worldwide problem. Up to 830,000 infants and toddlers die from accidental injuries every year. In addition to death cases, there are many sequelae due to damages caused by accidents. This caused decreased medical resources and increased the burden of financial pressure on parents. According to the reference, about 90% of accidents happen at home, so prevention strategies are necessary. There are many existing monitoring systems designed specifically for infants and toddlers, but most of the functions are relatively passive. Thus, the ability to reduce accidents is limited. In this paper, we proposed a baby accident prediction model based on recurrent neural networks, which is called BabyNet. First, to extract embedding features, each frame will send into three feature extraction modules: action module, object module, and pose module. The action module extracts two-scale features based on human detection. The object module gives each object an attention value and uses this value to fuse the object's features. The pose module processes the extracted skeleton feature, making it easier for the neural network to learn the relationship between pose and accident. Then, the probability of accidents is decoded by the RNN-based model. To consider the diverse predictability in accidents, the adaptive loss function is proposed, which estimates the instability of the video based on the optical flow. For evaluation and training, we built a dataset, baby video dataset, which contained about 1200 videos. According to the experimental results, BabyNet with adaptive loss function achieved 0.92 and 0.886 in AUC and AP. Also, BabyNet could predict accidents successfully before 3.6 seconds in advance.