Neuton's Transport type detection integration

[Vignesh R]

Hi,

I have integrated Neuton’s “Transport Type Detection” pre-trained model into my firmware, running on a nRF52840dk. For collecting 9-axis IMU data, I’m using the ST X-NUCLEO-IKS02A1 shield. Im using the standard example for iks02a1 shield to fetch data.

I’m successfully able to capture the 9-axis input for 64 iterations (window size :64)  and feed it to the Neuton model , and get predictions of some types. However, I need clarification on the following points:

Model Output Interpretation:
I receive numerical outputs 0 to 8 from the model's prediction result. I'm unsure what each target/class index corresponds to. Is there any documentation that I'm missing?

Sensor Configuration (Range & Output):
What are the expected sensor ranges and configurations (ODR, full-scale) and inference freq. to be used with this specific model?

Any guidance on interpreting model outputs and ensuring correct sensor setup will be very helpful.

Thanks in advance!

[Vidar Berg]

Hello,

Below is the response I received from our development team. I hope this helps.

For the “Transport Type Detection” preloaded use-case we have used the well known ‘Sussex-Huawei Locomotion Dataset’ (http://www.shl-dataset.org). This is a very extensive human movement research dataset with over 3000 hours of data collected by various motion sensors placed on different locations of the subject.
In relation to your questions

Model output interpretation: the preloaded model outputs 9 different labels from 0 to 8 which represent the following classes: null / still / walk / run / bike / car / bus / train / subway. However we cannot exactly map them at this point, because the original dataset consisted of 18 different labels with multiple versions of each class; e.g. still stand outside, still stand inside, still sit outside, etc. which were combined into 9 classes quite some time ago. If you are interested in this particular use-case implementation, we propose a meeting to discuss your application. We can retrain a transport type detection model that would suit your particular needs using this publicly available data or a custom dataset (https://neuton.ai/contactus).

Sensor Configuration: Data had been collected using the Huawei Mate 9 smartphone placed in 4 locations of the subject. please refer to the dataset collection readme on the SHL official website: http://www.shl-dataset.org/dataset/

Best regards,

Vidar