Distant Blurred Pedestrian Detection: Based on SSD Network with SRGAN Image Super Resolution

Traditional target detection methods have poor accuracy when processing lower resolution images with many pedestrians, particularly for small and blurry targets whose features are not manifested as significantly as the larger, clearer ones. This paper proposes an SSD pedestrian detection method with super-resolution image enhancement to address this problem. Because the small blurry pedestrians in the distance usually appear in the central part of the image, the model this paper proposed crops the central part of the screen and enhances it using SRGAN. After that, this paper use SSD to detect the enhanced image again. The detection results are fused with the normal detection results to obtain higher pedestrian detection accuracy. For the experiments, the AP of the proposed method reaches 0.693, and F1 comes to 0.709 in the street scene, which is 0.031 and 0.061 higher than the original SSD network. Meanwhile, the AP reaches 0.722 and F1 reaches 0.737 in the denser pedestrian scene, which is 0.044 and 0.085 more elevated than the original SSD network, reflecting the significant improvements in our proposed method's detection accuracy.