PLD_project : Pedestrain Light Detection Project for visually impared person

Introduction

The PLD project is Pedestrain Light Detection Project for visually impared person.

I train light-weight object detection models

TODO List

• [x] Add Yolo-v8 wrapper
• [ ] Add Yolo-v5 wrapper
• [ ] Add YoloR wrapper
• [ ] Add YoloX wrapper
• [ ] Add Demo notebook
• [ ] Add Demo python source code

Pedestrian light datasets

AI-Hub street walking dataset

DatasetDownloadLink

You need to join and request to download AI-Hub street walking dataset.

However, in the AI-Hub dataset, there is no distinction between traffic lights for car and pedestrians.

Therefore, i made an pseudo label using aspect ratio of bounding box.

If the bounding box is long left and right, it is classified as a car traffic light, and if it is long up and down, it is classified as a pedestrain light.

The generated pseudo label(classification) is available in

/label/AIHub/bbox_original : classname, x1, y1, x2, y2

/label/AIHub/bbox_normalized : classname, x, y, w, h (normalized)

class 0 : car traffic light

class 1 : pedestrain traffic light

There are no official split.

If you use our dataset split, Please refer to the split in the data/AIHub/labels/train and data/AIHub/labels/test in the directory and place the image files as follows.

bash data AIHub | images | | train | | | MP_KSC_000001.jpg | | | ... | | test | | | ZED1_KSC_003163_L_P000008.jpg | | | ... | labels | | train | | | MP_KSC_000001.txt | | | ... | | test | | | ZED1_KSC_003163_L_P000008.txt | | | ...

Test benchmark

model training

Now we have the pseudo labled traffic light dataset.

I finetune the off-the-shelf lightweight object detection models on collected traffic light dataset.

Model list

|Implemented|Model Name|Wrapper| |------|----|----| |:heavycheckmark:|yolo_v8|yolov8_wrapper| |:whitecheckmark:|yoloX|yolox_wrapper| |:whitecheckmark:|yoloR|yolor_wrapper| |:whitecheckmark:|yolo_v5|yolov5_wrapper|

Please check trainexp.sh or trainexp.bat

Model Zoo

All models are trained on GTX 980 ti

10 epoch, 20 epoch, 30 epoch models are seperately trained.

• AI-Hub train only

| Model | size
^{(pixels) | mAPtest
50-95
10 epoch | mAPtest
50-95
20 epoch |mAPtest
50-95
30 epoch | Speed
RTX-3090ti
(ms) | params
(M) | FLOPs
(B) | | ------------------------------------------------------------------------------------ | --------------------- | -------------------- | ------------------------------ | ----------------------------------- | ------------------ | ------------------ | ------------------ | | YOLOv8n | 640 | 0.2538 | 0.2818 | 0.2855 | | 3.2 | 8.7 | | YOLOv8s | 640 | 0.2893 | 0.3196 | 0.3241 | | 11.2 | 28.6 | | YOLOv8m | 640 | 0.3167 | 0.3597 | 0.3320 | | 25.9 | 78.9 | | YOLOv8l | 640 | 0.3260 | 0.3597 | 0.3452 | | 43.7 | 165.2 | | YOLOv8x | 640 | 0.3357 | 0.3640 | | | 68.2 | 257.8 |}

• mAP^test values are for single-model single-scale on AIHub dataset.

• AI-Hub + Imvisible train

| Model | size
^{(pixels) | mAPtest
50-95
10 epoch | mAPtest
50-95
20 epoch | Speed
RTX-3090ti
(ms) | params
(M) | FLOPs
(B) | | ------------------------------------------------------------------------------------ | --------------------- | -------------------- | ------------------------------ | ----------------------------------- | ------------------ | ------------------ | | YOLOv8n | 640 | | | | 3.2 | 8.7 | | YOLOv8s | 640 | | | | 11.2 | 28.6 | | YOLOv8m | 640 | | | | 25.9 | 78.9 | | YOLOv8l | 640 | | | | 43.7 | 165.2 | | YOLOv8x | 640 | | | | 68.2 | 257.8 |}

• mAP^test values are for single-model single-scale on AIHub dataset.

Color Classification Module

After detecting the pedestrain light, the color of pedestrain light should be classified.

Citation

If you find this repository helpful, please give us a citation.

@software{
    Hyungmin_PLD_porject_Pedestrain_2023,
    author = {Hyungmin, Kim},
    license = {GPL-3.0},
    month = {3},
    title = {{PLD porject: Pedestrain Light Detection project for visually impared person}},
    url = {https://github.com/khm159/PLD_project},
    version = {0.0},
    year = {2023}
}