Detection Of Vehicle Target Based On Deep Learning

The relation between people's growing travel demand and less space of transportation system has become the main contradiction in today's social transportation construction.Future cities should achieve harmonious development of time,space and sustainability concepts.Everyone has the right to enjoy convenient and efficient transportation services.So,The most indispensable part of history of human development is focusing on the survey of artificial intelligence and machine learning.These years,from the traditional target detection network to the convolutional neural network,from the two-stage target detection to the single-stage target detection,the detection accuracy at the algorithm level has been greatly improved.However,for the sub-scenario of vehicle detection,the detection rate and some very small targets on the real-time level are still difficult to detect,which also brings challenges to the field of actual transportation.Firstly,This paper analyzes the social significance and value of vehicle target detection research,then discusses the development and changes of vehicle target detection algorithm from artificial feature extraction to deep learning network in recent years.At the same time,The classic target detection algorithm based on deep learning around the world is explained in detail.Secondly,based on the YOLO(You Only Look Once)algorithm,this paper analyzes the neural network and detection principle of the algorithm by its high precision and fast speed.The real-time detection target to be realized in this paper requires higher speed in the algorithm.Considering the safety and complexity of the actual traffic conditions,it is required to detect not only the normal vehicle target,but also the scale-changed and very small targets in the distance.In this paper,feature extraction is used to extract features from very small targets.Furthermore,the K-means algorithm improved in this paper is used to cluster the vehicle target frames in the data set,besides anchor boxes are improved to the speed and accuracy of vehicle target detection.Finally,the end-to-end training of the deeplearning-based convolutional neural network is proposed in this paper.Our proposed deep learning vehicle detection algorithm achieves advanced performance in terms of accuracy and speed(up to 67 FPS)on MS COCO and PASCAL VOC benchmarks and new highway datasets,then compared with Retina,fast R-CNN,SSD,etc.Then,compared with the original algorithm,the proposed algorithm successfully tests the vehicles with small area and large scale changes that are not captured by the original algorithm.Since the vehicle has many positions and postures in three-dimensional space,the image information can not fully represent the spatial information.Therefore,the network structure,anchor frame design and loss function of the 3D YOLO network after upgrading the 2D YOLO network are expounded in this paper.At the same time,the algorithm is trained in the vehicle dataset to obtain the speed of the high algorithm and the accuracy under three different detection difficulties.The performance analysis is compared with the classic 3D target detection algorithm,and the detection is performed in the actual vehicle video.The real-time and accuracy are in line with the expected requirements for vehicle detection under actual scenarios.