Real-time Object Detection Method And Its Application In Traffic

Road obejct detection technology is a key technology in intelligent transportation.whose purpose is to detect the road targets according to the road monitoring image.But the road scene has various backgrounds and complicated targets.Today,road object detection is mainly done by Large computing server,but a set of road object detection servers is expensive.This paper designs a real-time road target detection embedded platform,which can realize real-time detection for bicycles,buses,cars,motorcycles and pedestrians for small application scenarios.The main research work of this paper are as follows:(1)A lightweight deep learning network suitable for embedded devices is proposed.Due to the existing YOLOv3 algorithm's detection speed is too slow and its model is too large.The backbone network and different spatial resolution feature models in YOLOv3 are optimized,and the real-time target detection MobileNetv1_yolov3lite_c2 network is designed.The computing speed of this network is 3.64 times higher than that of YOLOv3,and the model detection accuracy is reduced by about 10% compared with YOLOv3.(2)Optimization of the detection accuracy of the network.Based on LRM(loss rank mining)and label_smoothing methods,the loss function of MobileNetv1_yolov3lite_c2network algorithm is optimized and improved,and the mixup data augmentation method and cosine learning rate attenuation method are introduced to optimize the training process to improve the network detection accuracy of MobileNetv1_yolov3lite_c2.Simulation experiments show that the MobileNetv1_yolov3lite_c2 network has a 3.02% accuracy improvement,and can achieve a detection accuracy of 77.87% on the self-built data set.(3)Combining the hardware features of embedded device nvidia jetson TX2 and Mysql database,using multi-threaded design,realize embedded application of road target detection algorithm.The test results show that the MobileNetv1_yolov3lite_c2 network realizes real-time detection on TX2.This paper proposes the MobileNetv1_yolov3lite_c2 network,which can detect at26.67 FPS on the TX2 embedded board and achieve 77.87% detection accuracy on theself-built data set.