Vehicle Target Detection Based On Deep Learning

As the third wave of AI technology development with deep learning core technology has made breakthroughs in recent years and automated driving continues to advance towards commercialization,the requirements for target detection systems for automated vehicles are increasing.The autopilot engine of a driverless car has to make control decisions through environmental information gathered by sensors such as cameras and LIDAR.Therefore,there is a very high requirement for accuracy and real-time target detection in an automated driving scenario,as high latency in this scenario can have very serious consequences.The excellent performance of convolutional neural networks in image recognition is naturally introduced into the target detection task of computer vision.In recent years,the development of target detection algorithms has been pushing forward with increasing speed and accuracy.The development of target detection algorithms can be summarized into two lines: the Two-Stage algorithm represented by the R-CNN series and the One-Stage algorithm with YOLO as the pioneering work.The main work of this paper is to study vehicle target detection algorithms in the field of automated driving,and then propose a solution to optimize the performance of target detection algorithms by using lightweight convolutional neural networks to address the problem that previous target valuation models in automated driving scenarios do not meet the requirements of ground application in terms of efficiency.The efficiency problem is reflected in both the storage problem of the model and the detection speed problem.The first work in this paper is based on the YOLOv3 target detection algorithm,using the lightweight convolutional neural network MobileNet V2 to optimize its backbone network,a network structure MyMobileNet V2-YOLO V3 is redesigned based on the model efficiency and speed considerations,and a new loss function is redesigned to be more sensitive to small-scale predictions.deviations of the frame,and the dataset is augmented using a deep convolutional adversarial generation network taking into account the inherent defects of the vehicle dataset.The second work in this paper is to propose a vehicle target detection algorithm based on the lightweight convolutional neural network ShuffleNet.The basic unit is redesigned to optimize the network model for the problems of slow network speed caused by the large number of memory accesses,too many element-level operations and huge floating-point computations in the basic module of ShuffleNet.Finally,the experimental results of the two optimization algorithms proposed in this paper on the dataset are presented and compared with existing vehicle target detection algorithms,and it is concluded that the optimization of the vehicle target detection algorithm based on lightweight convolutional neural network improves the detection speed of the network without much difference in detection accuracy.