Vehicle Detection Based On Content Depth Multi-scale Convolutional Neural Network

Traditional object detection algorithms based on machine learning rely too much on prior knowledge,need to be designed manually,can only extract the surface features of the target.In contrast,deep learning algorithm is more advantageous in all aspects,which makes a breakthrough in the application of deep learning algorithm in vehicle target detection.The efficiency and accuracy of the detection method have been improved obviously,but it still can not meet the application requirements,and there are large and small inadaptability in the actual application scenarios,and a high precision and high efficiency algorithm is very important for the development of target detection.In view of the multi-scale problem of vehicles in the target detection task in the traditional CNN algorithm,the convolution feature was studied.It is common for vehicles in actual traffic images and videos to have large scale differences,and there are two key problems affecting their detection efficiency.On the one hand,the existing ROI pool destroys the structure of small scale vehicle targets;On the other hand,the representation capacity of a single network cannot support the detection of vehicles with large scale changes and large in-class distances.Therefore,this paper studies the multi-scale problem of vehicle input,and proposes a vehicle detection method based on content depth multi-scale convolutional neural network.Used for more rapid and accurate detection of vehicles with large scale changes.The main contributions of this paper are as follows:1.A new deconvolution feature pyramid network is introduced.Context information of CNN usually contains multi-level features.The deconvolution feature pyramid network can enhance the contextual content information of the high level because the low level feature graph has a large resolution,while the high level feature graph retains rich semantic information.2.A content-aware ROI pool was used to extract features for each candidate region to preserve the original structure of the small-scale target.Perception of ROI based on scale pool will deconvolution with bilinear nuclear application,to expand the feature of small candidate regions,thus to avoid repeating the value of the small target,using multilayer applied to CNN,then the pool will get elements in series with the low-level details and high-level semantic information fusion to detect targets,to maintain context information,Thus helping to maintain the original structural character.3.Introduce double branch decision network.The proposed vehicle detection method is divided into two parts according to the size of candidate regions,which can effectively reduce the problem of large changes in the target class.Thus,the detection accuracy of large and small targets can be improved.Then,the results predicted from multiple branches are integrated into the final detection results.It can accurately classify vehicles with large scale variations without adding additional computational costs.4.Vehicle experiments were carried out in public data sets KITTI and DETRAC.The results show that the accuracy of the proposed method is improved,and the false detection rate and time complexity are reduced.