Detecting pedestrians from photos using neural network with sliding window image pyramid
Matlab
Pedestrian Detection
Neural Network with Sliding Window Image Pyramid
Program to detect and classify pedestrians from an image. Achieved using 1 layer, 150 hidden unit neural network trained on the Daimler dataset which was border-reduced to improve training and decrease time and memory cost; the adjusted data is included here. After training the classifier, I used the sliding window image pyramid method to scan test images, where positive areas are marked with a bounding box as shown below. Currently my program will produce multiple bounding boxes in areas of interest, so the next step will be to merge them to produce a clear, precise detection result.
Challenges
I attempted to use a multiclass approach to train the network. However due to the massive range in possible negative examples (background variety) and not enough negative training data, a multiclass approach produced too many false positives even after feeding false positive test examples back into training data. That experiment can be found here. To overcome this, I then switched back to having a single output unit in the network, and used a very high prediction threshold (h >= 0.90 to 0.95) to increase precision and obtain the results shown.
