Research On Pedestrian Detection Technology On Rainy Days Based On Image Recognition

Pedestrian detection is widely used in vehicle infrastructure cooperative systems.Traditional pedestrian detection algorithms can achieve high accuracy and obtain reliable traffic data in sunny weather scenarios.However,when the vehicle is driving on rainy day,the detection accuracy will be greatly reduced because the rain presents multiple rain streaks near the camera to block the object in the field of vision,and the fog effect appears in the distance to block the sight.In view of the particularity of pedestrian detection on rainy days,the thesis proposed a pedestrian detection algorithm on rainy days,including de-raining module and pedestrian detection module,to improve the detection accuracy on different rainy days.The de-raining module used gaussian noise ratio and motion blur to evaluate the level of the rain condition,and find the areas that need to be focused in the input image through recurrent neural network,mainly the areas of rain streaks and their surrounding structures.Generative and discriminative Network is used to focus on these areas and de-rain.Compared with the classical de-raining algorithms in qualitative and quantitative aspects,the experimental results show that the de-raining module has a good de-raining effect in medium and heavy rain although its de-raining effect is slightly insufficient in light rain.The average PSNR and SSIM reach 28.31 and 0.838,respectively.The overall de-raining effect is better than the current advanced single image de-raining algorithms.Aiming at the problem of occlusion between pedestrians,the pedestrian detection module detected a pedestrian as a pair of keypoints—the top-left corner and bottom-right corner of the bounding box.The module introduced Corner Net-Lite algorithm and used the hourglass networks as the backbone network to extract image features.The output feature map is processed by corner pooling to obtain the position information of corners.We detected the corner position of pedestrians through two sets of heat layers and then optimize the loss of the embedded layer to make the distance between the two corners embedded into the same object smaller and smaller.At the same time,we calculated the loss of the offset layer,and adjusted the position of corners to generate a more compact prediction box.The proposed pedestrian detection algorithm is compared with the classical pedestrian detection algorithm in qualitative and quantitative aspects.The experimental results show that the average detection accuracy of the proposed algorithm is21.1%,48.1% and 60.9% higher than that of the YOLOv4 algorithm in light,medium and heavy rain scenarios,respectively.The pedestrian detection accuracy of the proposed algorithm is significantly better than that of the YOLOv4 algorithm on rainy days.In addition,pedestrian detection algorithm on rainy days in the thesis performs well on the real dataset,which can significantly improve the accuracy of pedestrian detection on rainy day scenes and have good robustness.