Research On Key Vehicle Detection Algorithms

With the development of artificial intelligence and computer vision,vehicle detection algorithms play an important role in intelligent traffic monitoring systems,intelligent transportation systems and smart cities.It is a prerequisite for realizing technologies such as vehicle path trajectory prediction,abnormal driving detection and crime tracking.At present,the vehicle target detection algorithm based on deep learning has been greatly developed.However,with the face of the camera angle change under the monitoring scene conditions,changing scale caused by the vehicle speed and requirement of the computing power by the convolutional neural network,challenging such as low detection speed and difficulty in embedded deployment still restrict the further development of this field.Therefore,based on the deep learning object detection algorithm,this paper studies the vehicle detection algorithm for detecting the location and basic model of the vehicle in the surveillance video scene,and proposes a series of improvement strategies based on the algorithm logic,performance and speed requirements of embedded deployment.The main contributions of this article are:(1)In order to deal with the vehicle scale variation issue and improve detection accuracy and speed,this paper proposes a vehicle detection algorithm based on multibranch perception enhancement network and original YOLO vehicle detection framework.The main idea is to use dilated convolution kernel with different dilated rates in the convolution layer to extract image features of different scales.First,the scale-aware strategy will cluster different vehicle scales in the training set and mark them as 3 different scale intervals.After that,the underlying backbone network will use a multi-scale dilated convolution module to extract features of the input image according to the scale interval.This module refers to the residual network structure,and different branches use the same network structure and calculation parameters to control the size of the image perception domain through different dilation rates,thereby improving training efficiency.Experiments show that the improved algorithm improves the detection accuracy by 6.5%,and improves the detection speed,effectively overcoming the scale variation issue.(2)In order to optimize the network structure to improve the algorithm detection speed and reduce the computing power requirements of the equipment,this paper proposes a new structured automatic pruning algorithm based on the original automatic pruning algorithm.First,for the deep convolutional neural network model of vehicle detection,heuristic search is used to quickly prune and evaluate the model to obtain hyperparameters that meet the performance.Then,according to the overall hyperparameter settings,the alternating multiplier optimization method is used to decompose the network pruning problem into sub-problems solved by gradient descent and Euclidean mapping.Finally,experiments show that the improved structured automatic algorithm can effectively optimize the pruning of deep-level convolutional neural networks,and can achieve a comprehensive compression rate of 1.8 times in the classification network.(3)In order to further improve the performance of the algorithm and achieve actual deployment,this article is based on NVIDIA’s Jetson TX2 deep learning development platform.After the hardware deployment of the improved vehicle detection algorithm,the improved pruning algorithm and GPU-driven acceleration framework are further optimized,and the final experiment verification algorithm meets the speed and performance requirements of embedded deployment,and a complete vehicle detection system is designed based on the detection algorithm to prove the feasibility of the actual embedded deployment of the deep learning vehicle detection algorithm.