Research On Dynamic Multiple Object Detection Model For Vehicle Forward Based On Deep Learning

The task of object detection is to find all interested objects in the image,and determine their location and category.In the process of driverless vehicle driving,detect all the potential dangerous objects rapidly and accurately in the surrounding complex traffic scenes is the basis for driving safety,and it is also a point to be solved by the current driverless technology.Combining the environment perception module in the test standard for China North Intelligent Vehicles City,this paper focuses on object detection technology in complex traffic scenes,takes deep learning as research method,constructs a dynamic mutiple object detection model for vehicle forward considering detection accuracy and speed.Main tasks are as follows:1.For the sake of exploring the mechanism of object detection based on deep learning and evaluate the object detection model scientifically,this paper analyzes the basic working principle of deep convolution network,combs the main evaluation indexes of the object detection model,and reasonably selects the evaluation indexes combining the research content and research points of the subject,and then defines mAP combine with FPS and Inference Time as the evaluation system of the model under the same test platform.2.In order to obtain the suitable data set,this paper analyzes and disassembles the current open source data set of object detection,extracts the applicable image data and its label files in the BDD100 K data set.The aim of expanding the object instance,this r paper sets up the acquisition platform,collects the image data of the test area in China North Intelligent Vehicles City with the real vehicle,annotates manually via LabImg.These two data sets are fused as the data sets of this paper.3.For the sake of constructing a multiple object detection model considering detection accuracy and speed,two-stage representative detection algorithm Faster R-CNN and one-stage representative detection algorithm YOLOv3 are compared and analyzed in this research.Determining YOLOv3 as the basic algorithm,and then training and testing the basic detection model.The result shows that mAP is 0.780,28 FPS and 30 ms Inference Time uder GPU performance,17 FPS and 256 ms Inference Time uder CPU performance.4.In order to optimize basic detection model under CPU performance,this paper takes lightweight model as the optimization direction,analyzes the common optimization methods of deep learning model.Replacing Darknet53 with MobileNetV2 as the backbone network of YOLOv3,replacing Adam optimization with Momentum optimization as optimizer of optimized model,and then on the basis of BN,adding GN to accelerate loss convergence.It can be found through traing and testing the optimized model that mAP is 0.800,61 FPS and 19 ms Inference Time uder GPU performance,44 FPS and 79 ms Inference Time uder CPU performance.Based on the improvement of 0.02 on mAP,the parameters of the optimized model are reduced by about 90% compared with the basic model,and the real-time detection can be realized under the CPU.5.For the sake of testing the performance of the detection model qualitatively,this paper designs two real vehicle test scenarios:China North Intelligent Vehicles City test and campus test.Through experiments,the dynamic multiple detection model designed and optimized by this paper institute has good robustness and applicability.