Falls, in particular falls among elderly, are a major public health problem. Automatic fall detection systems ensure that elderly people receive prompt assistance after experiencing a fall and can reduce the negative consequences the fall. The thesis focused on resolving problems related to the evaluation of fall detection algorithms on simulated datasets by conducting research on three specific issues. First, experiments were conducted to explore the influence of three variabilities in the experiment environment on the acquired acceleration data. The results revealed that all three variabilities cause heterogeneity between data in different fall detection datasets and based on these findings, a protocol for acquiring simulated falls with accelerometers was recommended. Next, a fall detection algorithm was designed and implemented in order to explore the influence of different window configurations in event-centered data segmentation on the performance of the detection algorithm. Based on the results, a method for implementing the data segmentation stage was proposed. Finally, a fall detection algorithm was trained with different numbers of subjects and trials to explore the influence on the performance outcomes. It was found that with more amount of data available for training, the performance follows an exponential raise and saturated at some point. Based on the saturation points, a recommendation for evaluating the relevance of existing fall detection datasets was made. The original scientific contribution of the thesis is: a dataset acquisition procedure for unbiased evaluation of fall detection algorithms, a method for selecting thresholds in fall detection algorithms and evaluation of relevance of the dataset for the efficacy of fall detection algorithms.