Underground video pedestrian detection method

Abstract

For problems of existing pedestrian detection methods based on deep learning such as large computation, detection efficiency relying on hardware performance heavily and so on, the pedestrian detection method based on SSD network was improved. A lightweight convolutional neural network based on DenseNet network is designed as basic network of SSD network to meet real-time detection requirements of underground video pedestrian, and an auxiliary network based on ResNet network is designed to strengthen feature extraction ability and improve correctness of pedestrian detection. An underground video pedestrian detection method based on the improved SSD network has been arranged in embedded platform Jetson TX2 for experiments. The results show that detection accuracy rate of the method for underground video pedestrian is 87.9% as well as nearly 100% for underground low-density pedestrian scene, and calculation speed achieves 48 frames per second, which is about 4.4 times as quick as the pedestrian detection method based on SSD network and meets real-time detection requirements of underground pedestrian.

Keywords

• underground pedestrian detection
• video monitoring
• deep learning
• ssd network
• convolutional neural network