Optimization of Convolutional Neural Networks for Real-Time Pedestrian Detection in Image Processing

Object detection is a very popular topic in image processing. Nevertheless, practical uses of CNNs for real-time tasks, especially pedestrian detection, still suffer from excessive computational and memory complexities. This work emphasizes the CNNs tuning process for increasing its efficiency in pedestrian recognition, especially under real-world (real-time) conditions. We changed the network architecture to reduce parameters, which still performs well. This can be done by decreasing the depth and width of the network and kernel size. Second, the low-level feature maps are thrown away to save memory consumption and important features for pedestrian detection. This is accomplished using a new feature selection algorithm that ranks feature maps in their importance to the detection task. In addition, for faster computation results, which are also accurate, the same research explores mixed-precision arithmetic-based methods in which some operations with less precision are done. Moreover, the impact of weight quantization on real-time performance is also studied. We evaluate the proposed optimizations over typical datasets and compare them to state-of-the-art techniques.