A Pedestrian Detection Method Based on YOLOv3 Model and Image Enhanced by Retinex

Pedestrian detection is a basic technology in the field of intelligent traffic video surveillance. It is also help for the optimization design of rail transport. It is known that the deep learning technology can achieve considerable performance on pedestrian detection. However, this kind of methods demand a large number of high-quality samples. In addition, the quality of data sample in the subway station is usually sensitive to the background environment, such as variant illumination or pedestrian density, which can significantly affect performance of the deep neural network. To solve this problem, this paper adopts an image enhancement policy based on the Retinex theory to preprocess training samples to reduce the influence of light changes. Firstly, we use the image enhancement method to enhance the contrast of the image and highlight the color of the object itself. Next, we put the initial sample into darknet frame with YOLOv3 to train the detection model1 and put the enhanced sample into the YOLOv3 to train the detection model 2. Finally, we tested these two models with 200 pedestrian pictures of four different scenarios. The experimental results show that the model trained by Retinex image enhancement has a more accurate detection rate of 94% compared with the model without the enhancement sample trained.