中国科学院自动化研究所机构知识库

With the development of pedestrian detection technologies, existing methods can not simultaneously satisfy high quality detection and fast calculation for practical applications. Therefore, the goal of our research is to balance of pedestrian detection in aspects of the accuracy and efficiency, then get a relatively better method compared with current advanced pedestrian detection algorithms. Inspired from recent outstanding multi-category objects detector SSD (Single Shot MultiBox Detector), we proposed a hierarchical convolution based pedestrians detection algorithm, which can provide competitive accuracy of pedestrian detection at real-time speed. In this work, we proposed a fully convolutional network where the features from lower layers are responsible for small-scale pedestrians and the higher layers are for large-scale, which will further improve the recall rate of pedestrians with different scales, especially for small-scale. Meanwhile, a novel prediction box with a single specific aspect ratio is designed to reduce the miss rate and accelerate the speed of pedestrian detection. Then, the original loss function of SSD is also optimized by eliminating interference of the classifier to more adapt pedestrian detection while also reduce the time complexity. Experimental results on Caltech Benchmark demonstrates that our proposed deep model can reach 11.88% average miss rate with the real-time level speed of 20 fps in pedestrian detection compared with current state-of-the-art methods, which can be the most suitable model for practical pedestrian detection applications.