Research On Vehicle And Pedestrian Recognition Algorithm Based On Improved YOLO

With the continuous development of deep learning technology,vehicle and pedestrian recognition applications are becoming increasingly widespread and have become an important branch in the field of computer vision.Traditional vehicle recognition methods are slow and have low accuracy.The accuracy of series algorithms based on YOLO has been improved,but there are also drawbacks such as large parameter quantities and low recognition speed.From the series algorithm of R-CNN to SSD and then to YOLO,each has its own advantages and disadvantages.YOLOv4 algorithm is a prominent algorithm in terms of recognition accuracy and speed.Considering the different requirements for algorithm recognition accuracy and speed in different scenarios,this thesis improves three vehicle and pedestrian recognition algorithms suitable for different work scenarios based on the YOLOv4 algorithm.First of all,there is still room for improvement in the m AP obtained by the YOLOv4 algorithm on the PASCAL VOC2007 dataset,which cannot meet the accuracy requirements of scenarios such as autonomous driving.In order to improve the recognition accuracy of YOLOv4 algorithm,the loss function and feature fusion network of the algorithm are improved.By replacing the CIo U loss function of YOLOv4 algorithm with SIo U,the boundary box regression of the model is more accurate.By strengthening the model’s fusion of shallow features,the model’s recognition ability for small targets has been improved.The experimental results show that the improved algorithm improves the m AP value by 2.2% on the VOC2007 dataset.Secondly,the vehicle recognition algorithm based on YOLOv4 has greatly improved its accuracy compared to traditional methods.However,due to its large computational parameters and deep network layers,there are problems such as slow recognition speed and high hardware requirements.In view of this,an improved YOLOv4 recognition algorithm is proposed.By introducing deep separable convolutions in the network to replace traditional convolutions,the redundant parameters and network layers of YOLOv4 are reduced,and they are combined into the resblock residual structure to maintain the transmission of gradient information.The experimental results show that the improved algorithm improves recognition speed by about 70% while ensuring high accuracy.Under the testing platform,the m AP on the VOC dataset was reduced by 3.2%,and the FPS was improved by 30.9 frames/s,resulting in good detection performance for small targets.Finally,considering that certain special scenarios have low requirements for recognition accuracy and the difficulty of target recognition in the scene is small,a more lightweight algorithm is improved.On the basis of applying deep separable convolutions,the channel number factor is introduced to make the convolutional network thinner.Simultaneously improve YOLO Head by using only two dimensional detection layers and simplifying the feature fusion network.The improved algorithm has good recognition accuracy and can recognize most vehicle and pedestrian targets by verifying the model with actual scene videos.The biggest advantage of this algorithm is that it relies on fewer hardware resources,can achieve real-time detection speed,and can meet some scenarios with low accuracy requirements and low computational power.