Research On Intelligent Traffic Vehicle Detection Method Based On Deep Learning

Vehicle detection is the foundation and core part of intelligent transportation system(ITS),which is of great significance to promote the construction of smart city and improve intelligent transportation system.The traditional vehicle detection algorithm uses a sliding window to traverse the whole image,which is easy to produce redundant windows and the speed is slow and it is difficult to accurately locate the target.In the vehicle detection task,due to the complex road traffic environment,the vehicle is easily affected by factors such as occlusion and scale changes,which leads to low detection accuracy of traditional vehicle detection methods.At the same time,vehicle images are mostly captured by road surveillance cameras,drones and other embedded devices as the carrier,while the traditional vehicle detection method has a large amount of calculation and high complexity,which is difficult to deploy on embedded platform with limited computing resources.With the development of deep learning theory,target detection algorithm based on convolutional neural network has stronger feature extraction ability and faster detection speeds.The application of lightweight target detection algorithms based on convolutional neural networks to vehicle detection tasks is in line with the development trend of highly intelligent and integrated intelligent transportation systems.DW-AD-YOLOv3 and G-YOLO target detection algorithms are designed to meet the requirements of vehicle detection task.DW-AD-YOLOv3 uses SoftPool to reduce the size of feature maps between densely connected units,which reduce the loss of fine-grained information in the pooling process.The attention mechanism is used to filter information that has a greater contribution to the final target prediction.The robustness of the model is improved by adding gaussian noise to the training sample for data enhancement.The improved dense connection structure with attention mechanism is used as the deep extraction layer of the network to obtain more semantic information.The deep separable convolution is used to reduce the number of parameters of the network and improve the detection speed.The experimental results show that,compared with the original algorithm,the average precision of the DW-AD-YOLOv3 algorithm is increased by 2.6 percentage points,and the number of parameters is reduced by 57%.In order to improve the generalization deployment ability of the algorithm in embedded devices,the structure of the cascade of Ghost convolution and attention mechanism is designed as the backbone network of G-YOLO.The attention mechanism is used to assign weights to some feature maps.All feature maps are obtained by performing simple linear operations on part of the feature maps,which reduces the amount of calculation.The receiving domain of the network is extended by adding the S-RFB module to the small target prediction branch.The experimental results show that,compared with the original algorithm,the average accuracy of the G-YOLO algorithm is increased by 2.5 percentage points,the model volume is only 32.9% of the original,and it has a higher detection speed.The DW-AD-YOLOv3 algorithm and G-YOLO algorithm proposed in this paper have high detection accuracy and detection speed with a lightweight network structure,which can meet the needs of vehicle detection tasks in complex road environments and have important significance for the development of intelligent transportation system.