Matlab GUI & implementation, training and testing of CNNs to detect pedestrians . Under 1s detection, 80% accuracy. (1 Sliding Window VGG) (2 HoG SVM VGG) (3 SSD)
CNN = Convolutional Neural Networks SSD = Single Shot (Multibox) Detector
General purpose pre-trained models (http://www.vlfeat.org/matconvnet/pretrained/) were trained (fine-tuned) for pedestrian detection.
The performance of each network is evaluated on the same set of 40 test images 640x480 gathered from various labelled machine-learning collections available online. Detection of pedestrians in those images have various level of difficulty due to different levels of occlusion, scale etc.
-
Sliding window at various scales submitted to CNN and evaluated with softmax using matconvnet-vgg-s 80% detection rate, averaging 22s per picture and 30% false positive rate
-
Sliding window HoG-SVM as region proposal. Regions submitted to CNN matconvnet-vgg-s for evaluation 90% detection rate, averaging 1s with 10% false positives
-
SSD using ssd- pascal-vggvd-300. 98% detection accuracy with an averaged time of 0.7s per image and 5% false positive rate
-
Pedestrian detector in videos using binary image transformations as a region proposal for the CNN matconvnet-vgg-s (1.)