The python tutorial over here should give you an idea on how to implement this.
The basic idea is for each window, compute the hog descriptors (using opencv, for example) and then multiply element-wise the HOG descriptors with the trained SVM weights (They should be the same size). After multiplying, add the bias (which is another output from the SVM classifier) to the previous result. If the result is positive, it's a positive match, otherwise it's a negative match.
Note: the sliding window size is the same size as the training images.
for each pixel in the image:
get a sub-image of size of sliding window
compute the HOG descriptors for this image
product = hog * SVM_weights //element-wise multiplication
response = product + bias
if response > 0:
print "no match"