Pedestrian Detection and Attribute Analysis Program Based on CNN

In recent years, deep learning object detectors including Fast/Faster R-CNN, SSD, R-FCN and Mask R-CNN have shown significant performance for general object detection except for pedestrians. The Region Proposal Network (RPN) in Faster R-CNN works well yet lacks of adaptability. Therefore, we propose an adaptive real-time pedestrian detection and attribute identification scheme based on Caffe. The first contribution is the adaptive threshold adjustment (ATA) algorithm for intelligent monitoring, utilizing the pedestrian movement information to adjust the threshold. Moreover, to overcome the time-consuming defect, we analyze the influences of the number of layers, the size of convolution kernels and the number of feature maps to reduce redundant computation while maintaining satisfactory performance. By optimizing the neutral network structure, choosing model parameters and data augmentation, a stable and well-performed model with fast detection rate and high accuracy is obtained. Besides, pedestrian information can also be identified in our program, offering better service in security monitoring, intelligent robots and other fields. Extensive experimental results demonstrate that even in complex and athletic scenarios, our method can make an improvement in quality and speed over state-of-the-art.