Vehicle Detection Based On Deep Learning In Natural Scenes

For driverless vehicles,detecting and identifying the existence of other vehicles around them is not only an important technology for their environment perception layer,but also a key function.Related sensors of driverless vehicles will transfer the sensor information to subsystem of the perception layer,in order to help the vehicles can use the data to estimate the environment and make the next decision.Because of the different natural scenes including cloudy,rainy,sunny,night and the situation of being obscured,the detection accuracy will be influenced.But in recent years,target detection based on Deep Learning can overcome the difficulties of traditional detection methods in real-time and accuracy and has achieved quickly detect and identify the target.In this paper,an algorithm based on deep learning is introduced to detect objections,and use it to detect vehicles in natural scenes.As follows is the main work:Consulted theories and literature relating to vehicles detection,then investigated and analyzed the current research status of target detection and video vehicle detection at home and abroad,found their advantages and disadvantages at the same time.Introduced the theories relating to deep learning,including the development of neural networks,the basic structure of convolutional neural networks,and some algorithmic principles involved in the training process of convolutional neural networks.Used HOG feature combining with support vector machine method and optical flow method to detect and identify vehicles,used image preprocessing technology for the purpose of eliminating the noise and image enhancement,which can get the high-quality features.Then used two traditional video vehicle detection methods to detect and identify the vehicles.At the last,by comparing traditional methods,and finding their problems,used the method based on deep learning named YOLO of this paper.analyzed relating technologies of YOLO V1,YOLO V2 and YOLO V3.Used collecting images in different natural scenes to make a vehicle dataset,built the deep learning framework(TensorFlow-Keras CPU),which used TensorFlow as backend,and Keras as headend.And installed relating dependent libraries that include OpenCV,Numpy,Matplotlib and so on;then embedded the improved YOLO V3 in the deep learning framework.Used the data set for model training,Then by iterative training,a model for vehicle detection and recognition was generated.Finally used this model to detect and distinguish the vehicles data that existed in natural scenes including sunny,rainy,cloudy,night,and target occlusion.and use this model to test and verify on UA-DETRAC data set,KITTI data set,and PASCAL VOC2007 data set;according to accuracy,missed detection,misdetection and real-time of the detection,comparing the original algorithm with algorithm of this paper,the algorithm of this paper significantly improved the phenomenon of missed detection and misdetection,and the detection accuracy was also significantly improved.