radar_obj_det_v2's Introduction

Instruction for Running TFLite Models on Raspberry Pi

1. Install the requirements.

pip install -U -r ./requirements.txt

2. Verify the label map.

Verify the labelmap.txt contains correct object classes.

3. Run the object detection live from camera using TFLite model.

python live_detection_tflite.py

Adjust the following variables:

flip: Set this to True or False depending on your camera orientation. Try this by experiment.
model_path: path to the tflite model to use.
input_size: Input size of selected tflite model:
model_type: type 'efficientdet' or 'retinanet' according to selected tflite model
det_score_threshold: detection confidence threshold. Based on my observation, for efficientdet: Try 0.6-0.7, for retinanet: Try 0.5-0.6.

Detection outputs

The following lines of codes in live_detection_tflite.py is the result of detection from each frame. Use it accordingly.

For Efficientdet

# Bounding boxes coordinates
bboxes = interpreter.get_tensor(interpreter_output[1]['index'])[0]
# Detected objects class ID
class_id = interpreter.get_tensor(interpreter_output[3]['index'])[0]
# Detection scores
scores = interpreter.get_tensor(interpreter_output[0]['index'])[0]

For Retinanet

# Bounding boxes coordinates
bboxes = interpreter.get_tensor(interpreter_output[1]['index'])[0]
# Detected objects class ID
class_id = interpreter.get_tensor(interpreter_output[2]['index'])[0]
# Detection scores
scores = interpreter.get_tensor(interpreter_output[3]['index'])[0]

4. Run the object detection live from camera using YOLO model.

python live_detection_yolo.py

Adjust the following variables:

flip: Set this to True or False depending on your camera orientation. Try this by experiment.
model_path: path to the tflite model to use.
input_size: Input size of selected tflite model:
det_score_threshold: detection confidence threshold.

Recommend Projects