As people grow older their motor functionality tends to deteriorate. This deterioration often implicates a slower, more careful walk and the use of walking aids. One way to monitor a person’s health is by observing a person’s daily physical activity. This daily activity can be measured with the help of programs in wearable devices, such as a smartwatch step counter. Most conventional step counting algorithms aim to count steps from healthy humans with a standard gait. But because of the lower amplitude of the signals collected from an elderly person with deteriorated gait, many of these step counting algorithms will struggle to detect walk. At the same time, monitoring the activity level and walk is arguably even more important in elderly people, as it can be used in both preventive and diagnostic measures regarding the person’s health.In this thesis, an algorithm was developed that counted steps and classified accelerometer data into the four different types; no walk, normal walk, slow walk and walk with aid. The algorithm did this with the help of machine learning, signal processing methods and a dynamic threshold. The thesis aimed to answer how suitable acombination of machine learning and dynamic thresholding could be for the purpose of walk detection in elderly, and what the accuracy of the developed algorithm was. The result showed that the walk classification had an accuracy of 86.4%, while the step counting had an accuracy of 88%. Ultimately, it was concluded that more data was needed to properly evaluate and train the algorithm, but that it held protentional for future work to be done.