YOLOv5-Tracker

This repository contains extra methods on top of the ultralytics yolov5 repository Extra features are:

Light, strong custom object tracker
Key point finder and feature matcher with optical flow
Variance of the image laplacian to calculate image blur(To avoid possible false detections)

Statistical Accuracy

Object Detection Accuracy

Tracker Algorithm Accuracy

Whole pipeline is runned on 100 randomly selected videos from Berlin. 
Within each video, a json file that indicates the individual car locations is also available. (private data)
Model output is compared with ground truth data.
Tracker algorithm success calculated as 89.5% in the test.

Due to some privacy reason, the geo location of the videos are not to be shared. But an example output of map module(not included in the repository) is represented here:

Requirements

To install the pipeline run:

$ pip install -r requirements.txt

Inference

To run the inference on test video simply run:

$ python analysis/main.py

To run the inference on a defined video simply run:

$ python analysis/main.py --source /path/to/video.mp4

In default script saves the video output, if you do not want to save them, simply assign video_output argument to False by:

$ python analysis/main.py --source /path/to/video.mp4 --video_output False

If you want a faster inference time, you can reduce the image size by:

$ python analysis/main.py --source /path/to/video.mp4 --img_size 320

python analysis/main.py --source /path/to/video.mp4 --img_size 160

Update on 28th January

Car plate number detection & Anonymisation script is added.

Update on 30th January

Face detection, pistol detection, human face & car plate(both in one) detection models are trained and added in to google cloud

If you want to use any of these models, please first download the trained model then run:

$ python analysis/main.py --source /path/to/video.mp4 --weights path/to/downloaded/model.pt

or if you want to anonymized objects;

$ python analysis/anonymization.py --source /path/to/video.mp4 --weights path/to/downloaded/model.pt

Models have trained with Nvidia-Tesla V100 GPU over 100 epochs on the private datasets.

Model	size	AP₅₀
YOLOv5s_carPlate	640	92.3
YOLOv5x_carPlate	640	97.5
YOLOv5s_pistol	640	95.4
YOLOv5s_face	640	98.2
YOLOv5s_faceAndCarPlate	640	98.2

YOLOv5s_carPlate in real-time:

CoskunGorkem/YOLOv5-Tracker