Collecting hands data is important for many cognitive studies, especially for senior participants who has no IT background. For example, alternating hand movements and imitation of gestures are formal cognitive assessment in the early detection of dementia. During data collection process, one of the key steps is to detect whether the participants is following the instruction correctly to do the correct gestures. Meanwhile, re-searchers found a lot of problems in TAS Test hand movement data collection process, where is challenging to detect similar gestures and guarantee the quality of the collect-ed images. We have implemented a hand gesture detector to detect the gestures per-formed in the hand movement tests, which enables us to monitor if the participants are following the instructions correctly. In this research, we have processed 20,000 images collected from TAS Test and labelled 6,450 images to detect different hand poses in the hand movement tests. This paper has the following three contributions. Firstly, we compared the performance of different network structures for hand poses detection. Secondly, we introduced a transformer block in the state of art network and increased the classification performance of the similar gestures. Thirdly, we have created two datasets and included 20 percent of blurred images in the dataset to investigate how different network structures were impacted by noisy data, then we proposed a novel net-work to increase the detection accuracy to mediate the influence of the noisy data.