Our project addressed the question of whether using multiple IMUs along a specific walking path would improve position accuracy.
To compute the most accurate position possible we used the Pedestrian Dead Reckoning (PDR) method. PDR calculates the current position based on previously determined positioning, using an estimation of speed and heading direction over a short time period.
Initially, we calculated the distance traveled along each step using classic methods, but switched to deep learning after receiving unsatisfactory results, and more specifically the transfer learning method. As we progressed, we tried a few complicated networks like ResNet50 before moving to SimpleNet, a simpler network with fewer parameters.
For the training dataset, we recorded a total of 10 walking routes lasting 2 minutes each. In order to track information about the walk, we used a hardware component named MIMU4444 consisting of four IMUs each equipped with three acceleration sensors and three gyroscopes, and another component named MRU-PD to record the exact location. We used the information collected by both devices to train our learning model. Additionally, we used the publicly available dataset RIDI [2] that contains multiple recordings of walking routes in order to create a rich and diverse dataset.
The results obtained from our extensive experimental evaluation support the value of using a number of IMUs instead of just one, and they demonstrate the high precision in position that can be achieved by increasing the number of IMUs.