Vehicle Detection And Recognition In Traffic Scenes Based On Deep Learning

With the rapid growth of vehicles,traffic congestion is increasingly severe.Intelligent transportation systems have aroused lots of interests due to its significant contribution in providing an efficient and safe traffic environment.Vehicle detection and recognition is one of the key issues in intelligent transportation systems.Great progress has been made in deep learning based vehicle detection and recognition.However,the performance of vehicle detection and recognition in traffic environment is still unsatisfactory because of variability of vehicle size,cluttered scenes,lighting variation and occlusion.Against this background,lots of research on deep learning based vehicle detection and recognition are done in this article,therefore,we propose Separable Reverse Connected network(SRC)for efficient multi-scale vehicle detection and recognition.Main works could be concluded as follows:1.Analyzes of the deep learning detection frameworks are made in this article.The performance of different feature extractors and detection networks shall be learned by means of comparative experiments.The structure with better performance in vehicle detection and recognition is selected as the basic model.2.Based on the results of comparative experiments,we propose a multi-scale vehicle detection algorithm called Separable Reverse Connected network(SRC).Separable convolution network is introduced for sparse representation of heavy feature maps generated from feature extractors,while reverse connected network improves the ability of multi-scale feature representation by enriching the semantic context information of previous layers.3.A series of optimization algorithms are introduced in order to improve the performance of our models.Multi-scale training and Online Hard Example Mining(OHEM)are applied for training optimization,while model compression reduces the quantity of parameters and calculation for network efficiency.In order to evaluate our network models,experiments are carried out on the public datasets Pascal VOC and MS COCO.The experimental results indicate that SRC networks achieve better results than state-of-the-art detection frameworks such as FPN and YOLOv2.Additionally,optimization algorithms further enhance the performance of the network,which leads to real-time detection with the ability of multi-scale vehicle detection and multi-category vehicle recognition.