This is a very simple demo of vehicle detection,I use JavaFx to make a very simple GUI.
When you're using this gui to detect vehicle video or images, it's actually calling darknet.exe by command line.
I think you would like darknet! You can get more information about it in here (Original) or here (AlexeyAB) ,the darknet.exe of this repository compiled by vs2015 and use src of AlexeyAB's repository.
My environment is CUDA 9.1 with cuDNN 7.0.And OpenCV 3.0 .You can download the *.dll in here with the password: ijek
If you dont have NV GPU, you can use darknet_no_gpu.exe rename to darknet.exe
We can use darknet to train detector easyly with two amazing detection methods：Yolo v2 and Yolo v3.
Yolo v2:
Yolo v3:
I use Pre-training model to train four detector on UA-DETRACT training set.
1.single class Yolo v2 detector
2.four-classes Yolo v2 detector
3.single class Yolo v3 detector
4.four-classes Yolo v3 detector
Single class detector only detect car,but four-classes detector detect four kinds of car includes car,van,bus,others.You can download these in here with the password: ijek
UADETRAC-TrainSet：
| car | van | bus | others | |
|---|---|---|---|---|
| count | 503853(84.21%) | 57051（9.53%） | 33651（5.62%） | 3726（0.62%） |
The model I trained is not good enough.This is PR curve of my detector with others detectors at UA-DETRAC website:
Yolo is very Fast!Yolo v2 detector achive 54 FPS on GTX1070,and Yolo v3 detector achive 29.4 FPS because it uses a deeper network to achive higher mAP on test benchmark.And Yolo v2 and v3 detector can achive 16 an 6.9 FPS separately on GTX860m.
Train Sample
Detect Simple(Yolo v2 four-classes)：
