The data features that you choose to use are relatively independent of the classification technique chosen. Read what data features other researchers have used when classifying activities of daily living.

The biggest advantage of deep learning over conventional supervised methods is its ability to learn features automatically. This means you do not need to design features by yourself, and just take the raw accelerometer data or processed data (e.g., ZCA proprocessing, FFT) as input to the first layer of the deep learning method you choose. 

Depends on how you are planning to model the subject. You can try going for a typical sliding window approach and use a deep learning approach for classification, or maybe go for a time-step solution using RNN LSTM. These models might be simple but it will be a good starting point and baseline for other advance methods that you will try in the future.