Fast Pedestrian Detection in Surveillance Video Based on Soft Target Training of Shallow Random Forest

In recent years, deep learning algorithms have achieved top performances in object detection tasks. However, in real-time, systems having memory or computing limitations very wide and deep networks with numerous parameters constitute a major obstacle. In this paper, we propose a fast method for detecting pedestrians in surveillance systems having limited memory and processing units. Our proposed method applies a model compression technique based on a teacher-student framework to a random forest (RF) classifier instead of a wide and deep network because a compressed deep network still demands a large memory for a large amount of parameters and processing resources for multiplication. The first objective of the proposed compression method is to train a student shallow RF (S-RF), which can mimic the teacher RF's performance, by using a softened version of the teacher RF's output. Second, a deep network cannot easily detect small and closely located pedestrians in a surveillance video captured from a high perspective because of frequent convolutions and pooling processes. In this paper, adaptive image scaling and region of interest with S-RF were therefore combined to allow fast and accurate pedestrian detection in a low-specification surveillance system. In experiments, our proposed method achieved up to a 2.2 times faster speed and a 2.68 times higher compression rate than teacher RF and a better detection performance than several state-of-the-art methods on the Performance Evaluation of Tracking and Surveillance 2006, Town Centre, and Caltech benchmark datasets.