7.9.17. PIR Detection

Edge AI solution for PIR (passive infrared) sensor based motion classification on the MSPM0G5187 with NPU acceleration, also available for the SimpleLink CC13xx/CC27xx/CC35X1 wireless MCUs.

Example folder: examples/pir_detection (GitHub)

7.9.17.1. Overview

Traditional PIR sensors only detect presence or absence of motion and cannot tell its source. Pets and environmental factors cause false alarms, and manual threshold tuning is unreliable across environments. This example classifies PIR sensor signals into motion categories, which reduces false positives while keeping detection sensitivity.

Application: security systems, smart home automation, occupancy sensing
Task: classification (task_type: pir_detection), 3 classes
Primary device: MSPM0G5187
Model: PIRDetection_model_1_t
Targets of the reference solution: real-time classification, more than 90% accuracy, low power

7.9.17.2. System Components

Hardware: MSPM0G5187 with integrated NPU, and the EdgeAI Sensor Boosterpack (TIDA-010997) with a PIR sensor
Software: Code Composer Studio 12.x or later, MSPM0 SDK 2.08.00 or later

7.9.17.3. Available Configurations

Config file	Device	Feature extraction preset / epochs
`config_MSPM0.yaml`	MSPM0G5187	`PIRDetection_125Input_25Feature_25Frame_1InputChannel_2D_FixedPoint`, 100 epochs
`config_CC1312.yaml`	CC1312	`PIRDetection_125Input_25Feature_25Frame_1InputChannel_2D`, 50 epochs
`config_CC1314.yaml`	CC1314	same as CC1312
`config_CC1352.yaml`	CC1352	same as CC1312
`config_CC1354.yaml`	CC1354	same as CC1312
`config_CC2755.yaml`	CC2755	same as CC1312
`config_CC35X1.yaml`	CC35X1	same as CC1312

The MSPM0 config uses the fixed-point feature extraction that matches the integer pipeline on the device; the other configs use the floating point version.

7.9.17.4. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

# MSPM0G5187 (primary)
./run_tinyml_modelzoo.sh examples/pir_detection/config_MSPM0.yaml
# other devices
./run_tinyml_modelzoo.sh examples/pir_detection/config_CC1352.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\pir_detection\config_MSPM0.yaml

7.9.17.5. Dataset

pir_detection_classification (https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/pir_detection_classification_dsk.zip) contains PIR recordings captured with the EdgeAI Sensor Boosterpack from a range of 6 to 8 meters:

Class	Description
Human motion	Movement patterns characteristic of humans
Background motion	Environmental disturbances and noise
Dog motion	Movement patterns characteristic of dogs

7.9.17.6. Feature Extraction Pipeline

The PIRDetection_125Input_25Feature_25Frame_1InputChannel_2D presets process the raw signal as follows:

ADC sampling
DC offset removal
Windowed processing
Symmetric mirroring
FFT computation
Magnitude calculation
Average pooling
Additional features: zero crossing rate (ZCR), slope changes, dominant frequency
Concatenation of all features into the model input

7.9.17.7. Configuration (`config_MSPM0.yaml`)

common:
  task_type: pir_detection
  target_device: MSPM0G5187
dataset:
  dataset_name: pir_detection_classification
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/pir_detection_classification_dsk.zip
data_processing_feature_extraction:
  data_proc_transforms: []
  feature_extraction_name: PIRDetection_125Input_25Feature_25Frame_1InputChannel_2D_FixedPoint
  variables: 1
training:
  model_name: PIRDetection_model_1_t
  batch_size: 64
  training_epochs: 100

PIRDetection_model_1_t is a compact CNN with about 53K or more parameters, NPU compatible and optimized for multivariate input signals. The MSPM0 config trains with batch size 64, learning rate 0.00001, weight decay 1e-20 and 100 epochs; the CC13xx/CC27xx/CC35X1 configs use batch size 64, learning rate 0.001 and 50 epochs. The optimizer is the default SGD. Compilation uses the TI Neural Network Compiler and generates model.a and tvmgen_default.h.

7.9.17.8. Results

The example readme.md reports about 98% accuracy for CC1352 with floating point feature extraction and about 92.46% for MSPM0 with fixed point feature extraction (earlier release). In the automated 1.5.0 config test run:

Config	Test accuracy	AUC ROC
`config_CC1312`, `CC1314`, `CC1352`, `CC1354`, `CC2755`, `CC35X1`	94.81%	0.826
`config_MSPM0.yaml`	65.40%	0.795

Note

The MSPM0 test-run accuracy (65.40%) is far below the 92.46% the example readme quotes from an earlier release. The reasons were not investigated here; check the printed confusion matrix of your own run and consider tuning epochs and learning rate for the fixed-point configuration.