Accurate and reliable automatic fall detection based on wearable devices enables elderly people to receive instant treatment and can alleviate the severe consequences of falls. Falls are abnormal activities that occur rarely compared with normal daily activities; therefore, fall detection can be considered a one-class classification problem. However, it is difficult and dangerous to collect sufficient fall data in practice, making it difficult to use supervised learning methods to detect falls automatically. Among wearable devices, wrist-worn devices for fall detection are more likely to be accepted because of comfort; however, high accuracy cannot typically be realized due to sensitivity to interference from the diverse activities of the hand and wrist. Combining ensemble stacked autoencoders (ESAEs) with one-class classification based on the convex hull (OCCCH), this paper proposes a novel intelligent fall detection method, namely, ESAEs-OCCCH, which is based on accelerometer data from a wrist-worn smart watch. In the proposed method, ESAEs are first adopted for unsupervised feature extraction to overcome the disadvantages of artificial feature extraction, namely, the requirements in terms of experience and time. Then, OCCCH is used for pattern recognition. Finally, the majority voting strategy and weight adaptive adjustment strategy are combined to improve the performance and stability of fall detection. According to the behavioral characteristics of the elderly, the uncertainty of activities of daily living (ADLs) and fall activities (FAs), and the influence of the intense activities of the hand on the accelerometer signal, thirteen FAs and sixteen ADLs, including intense hand and wrist activities, are simulated by young volunteers of various genders, ages, heights and weights in two experiments. The experimental results demonstrate the performance and stability of the proposed method.