2

I am trying to train cascade to detect motorcycle. I tried LBP (with stage 15) and it's giving too much false negatives, Here is what I did till now,

I recorded required traffic videos . Developed a program to track all moving objects and when it crosses line of detection, cv::imwrite() to save the Rectangular ROI. Manually separated which are Motor Cycles and which are not, cropped them and made sure that the others do not have any motorcycles. Like this I have collected around 257 positive images and 653 negative images (I can get more negatives if required),

If you would like to have a look at the collected data, I have uploaded them here :

It looks like this :

positives : (yes, they are of different sizes)

enter image description here , enter image description here , enter image description here , enter image description here, enter image description here , enter image description here

negatives are cars, trucks, people, etc...

I generated LBP cascade using opencv_traincascade with numStages 15 and 20 . They are giving too much (around 55%) false negatives

I tried to generate HOG cascade using https://github.com/opencv/opencv/blob/master/samples/cpp/train_HOG.cpp , but it requires be to provide samples of fixed sizes , that too 64x128 (which will totally deform my images, I don't think that's going to work) .

By seeing the Images above, could you please suggest some ideas how I achieve my requirement ? Should I try HOG ? HAAR ? LBP ? should I increase/decrease the numStages ? how to decrease false negatives ?

Your inputs are most valued, Thanks.

Severus Tux
  • 265
  • 3
  • 13

1 Answers1

1

You may try TensorFlow. I have applied it for vehicle classification. Many pictures are needed by to teach network. But results were not so good. For any camera positions need to train many times.

I am resolving the same problem now. What algorithm is better and will be more suitable for solving number plates searching problem on an image? I've already solved the char recognition task. I have read about Haar, HOG, LBP. Does anything else (probably more effective than these algorithms) exists?

I have attached the input image example. The image has compressed for 5 fold. For example, in this image, I draw a red rectangle around the number plate. the real road photo

Its the worst case, usually number plate and image have better quality, contrast, and brightness. This photo was made on a winter day when the weather was dirty and cloudy.

Accuracy is more important than performance and speed of the algorithm, I have a good enough CPU.

Orlov Const
  • 332
  • 3
  • 10