IGN is acronym of Ignatov, and is a convolutional neural network (CNN) based model for performing the human activity recognition (HAR) task based on the 3D accelerometer data. In this work we use a modified version of the IGN model presented in the paper[2]. It uses the 3D raw data with gravity rotation and supression filter as preprocessing. This is a light model with very small foot prints in terms of FLASH and RAM as well as computational requirements.
This network supports any input size greater than (20 x 3 x 1) but we recommend to use at least (24 x 3 x 1), i.e. a window length of 24 samples. In this folder we provide IGN models trained with two different window lenghts [24 and 48].
The only input required to the model is the input shape, dropout ratio, and the number of output classes.
In this folder you will find multiple copies of the IGN model pretrained on a public dataset (WISDM) and a custom dataset collected by ST (mobility_v1).
| Network Information | Value |
|---|---|
| Framework | TensorFlow |
| Params | 3,064 |
For an input resolution of wl x 3 x 1 and P classes
| Input Shape | Description |
|---|---|
| (1, wl, 3, 1) | Single ( wl x 3 x 1 ) matrix of accelerometer values, wl is window lenght, for 3 axes and 1 is channel in FLOAT32. |
| Output Shape | Description |
|---|---|
| (1, P) | Per-class confidence for P classes in FLOAT32 |
| Platform | Supported | Recommended |
|---|---|---|
| STM32L4 | [x] | [] |
| STM32U5 | [x] | [x] |
Measures are done with STM32Cube.AI Dev Cloud version 10.0.0 with enabled input/output allocated options and balanced optimization. The inference time is reported is calculated using STM32Cube.AI version 10.0.0, on STM32 board B-U585I-IOT02A running at Frequency of 160 MHz.
Reference memory footprint and inference times for IGN models are given in the table below. The accuracies are provided in the sections after for two datasets.
| Model | Format | Input Shape | Series | Activation RAM (KiB) | Runtime RAM (KiB) | Weights Flash (KiB) | Code Flash (KiB) | Total RAM (KiB) | Total Flash (KiB) | Inference Time (msec) | STM32Cube.AI version |
|---|---|---|---|---|---|---|---|---|---|---|---|
| IGN wl 24 | FLOAT32 | 24 x 3 x 1 | STM32U5 | 2.03 | 1.91 | 11.97 | 13.61 | 3.94 | 25.58 | 2.25 | 10.0.0 |
| IGN wl 48 | FLOAT32 | 48 x 3 x 1 | STM32U5 | 4.56 | 1.91 | 38.97 | 13.61 | 6.47 | 52.58 | 8.17 | 10.0.0 |
Dataset details: A custom dataset and not publically available, Number of classes: 5 [Stationary, Walking, Jogging, Biking, Vehicle]. (We kept only 4, [Stationary, Walking, Jogging, Biking]) and removed Driving, Number of input frames: 81,151 (for wl = 24), and 40,575 for (wl = 48).
| Model | Format | Resolution | Accuracy (%) |
|---|---|---|---|
| IGN wl 24 | FLOAT32 | 24 x 3 x 1 | 94.64 |
| IGN wl 48 | FLOAT32 | 48 x 3 x 1 | 95.01 |
Confusion matrix for IGN wl 24 with Float32 weights for mobility_v1 dataset is given below.
Dataset details: link , License CC BY 2.0 , Quotation[1] , Number of classes: 4 (we are combining [Upstairs and Downstairs into Stairs] and [Standing and Sitting into Stationary]), Number of samples: 45,579 (at wl = 24), and 22,880 (at wl = 48).
| Model | Format | Resolution | Accuracy (%) |
|---|---|---|---|
| IGN wl 24 | FLOAT32 | 24 x 3 x 1 | 91.7 |
| IGN wl 48 | FLOAT32 | 48 x 3 x 1 | 93.67 |
Please refer to the stm32ai-modelzoo-services GitHub here
[1] “WISDM : Human activity recognition datasets". [Online]. Available: "https://www.cis.fordham.edu/wisdm/dataset.php".
[2] “Real-time human activity recognition from accelerometer data using Convolutional Neural Networks, Andrey Ignatove". [Online]. Available: "https://www.sciencedirect.com/science/article/abs/pii/S1568494617305665?via%3Dihub".