Improved Real-time Pedestrian Detection Method

With the development of deep learning, algorithm research in object detection has also made significant progress. At present, there are many excellent object detectors, such as Faster R-CNN, SSD, etc. YOLO is one of the most outstanding detectors and has many practical applications in industry. Based on this, we study how to apply it to real-time pedestrian detection, and improve the detection performance. Based on the YOLO algorithm, we mainly carry out the following work: 1). Apply parameterized ReLU as main activation function of network; The classification subnet is separated from the regression subnet, and the parameters are no longer shared between the two subnetworks to improve the network performance; 2). Design loss function to reduce the effect of foreground-background class imbalance caused by anchors, and the classification is improved by hard sample mining. The weight coefficient is designed to improve the localization of the network for small objects. 3). Design the parameters of the anchors to optimize the localization of pedestrian objects.