I want to do pedestrian detection for a real time application. But the platform that I am using does not have the option for training feature vectors. ( that is I cannot train feature vectors). I only have the option of obtaining HOG feature vector. After this I want to do human detection.
So, can someone help in doing this classification without machine learning algorithm.
(The processor and remaining hardware specification have to be decided only after finalizing the classification part.
I also have the option of using FPGA)
To "know" what a pedestrian looks like, a detection algorithm needs to learn. To avoid the hassle of training, you could use some pre-trained classifier, and see if it is effective for your needs. Best results will be obtained by training a classifier on data that is representative of your test application. You can always train your system on a PC and push it onto an FPGA. The attached paper should be a good starting point, if you haven't implemented it already. Their pre-trained HOG-based detectors may be found in publicly accessible libraries, such as, OpenCV.
Article Histograms of Oriented Gradients for Human Detection
You can use clustering techniques instead of classification techniques.
Think of the world "learning" as "modeling". Your program needs a model that discriminates whether a certain region in the image is a human or not. There are very good reasons why machine learning is so popular nowadays, above all, because it is very difficult to make such a model by "hand".
I think there is a misunderstanding in how you use the term "feature vectors". You don't train feature vectors. A feature vector is a representation of an observation of "something". For example, the HOG values of a region in an image could form part of the feature vector that describes that region.
Machine learning algorithms (e.g. a classifier) are trained (i.e. a model is constructed and parametrized) upon a collection of observations (i.e. feature vectors) of the phenomenon of interest (e.g. pedestrians). In some cases, the model also needs observations of "non-pedestrians" so that he can tell the difference, but that depends on the method you are using.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification