The Research Of Object Detection Algorithm In Traffic Scene Based On Deep Learning

In recent years,driven by computer,communication,artificial intelligence and other technologies,the development of autonomous driving and ADAS technology is in full swing,which will play a more important role in solving various traffic problems.As the key link of autonomous driving and ADAS,environmental perception is the precondition and foundation of the whole system's safety and robustness.Traffic scene object detection based on computer vision technology,as an important task of autonomous driving environment perception,requires simultaneous detection of vehicles,pedestrians and other targets,and faces many challenges such as multiple target scales,diverse attitude and angles,occlusion,complex environment background,changing illumination intensity,high accuracy and high detection speed.Compared with traditional algorithms based on manual design features and sliding windows,object detection algorithm based on deep learning has many advantages,such as high accuracy,versatility,strong task migration ability,low cost of engineering development,optimization and maintenance.It can achieve the effect that traditional methods can not achieve and become an effective solution to overcome the difficulties mentioned above in the task of object detection in traffic scenes.This paper improves and optimizes the YOLOv3 object detection algorithm based on deep learning.Aiming at the problem of large parameters and computation of original YOLOv3 algorithm,a lightweight MobileNetV2 network based on deep separable convolution and reverse residual blocks is used to improve the backbone network of object detection algorithm,which reduces the amount of parameters by 58.3% and improves the detection speed by 47% in the case of slight loss of detection accuracy.The k-means algorithm is used to cluster the ground turth bounding box information of the target in the KITTI dataset.The optimized anchor boxes size is more in line with the characteristics of the target in the traffic scene,and the detection accuracy is improved by 1.35%.In addition,the effects of different input sizes and the number of anchor boxes on the performance of object detection algorithm are experimentally explored and analyzed.Considering the detection speed and accuracy,the most balanced input sizes and the number of anchor boxes are selected.In addition,the GIoU loss function is adopted as the bounding box regression loss function of YOLOv3 algorithm,which further improves the object detection accuracy in traffic scene.Finally,the improved YOLOv3 object detection algorithm in this paper achieves 82.29% detection accuracy and 35.7 FPS detection speed on test set.It achieves high detection accuracy and fully meets the requirements of real-time detection.The generalization ability of the algorithm is tested on the actual road images of Changsha,which proves the effectiveness of the algorithm in this paper.