Research On Vehicle Object Detection Algorithm Based On Heat Map And Weighted Feature

With the increasing development of deep learning and computer technology,the vehicle object detection algorithm based on deep learning has a great improvement in speed and accuracy compared with traditional methods.However,the current vehicle object detection algorithm is basically based on the border detection algorithm,which has problems such as complex algorithm,high computational resources and magnitude parameters,and there is still a lot of room for improvement.Therefore,how to simplify the framework of the algorithm,reduce the number of parameters and eliminate redundant calculation has become a challenging work.In this paper,the vehicle object detection method and weighted feature fusion method based on heat map are studied,and the deep aggregation network DLA-34 is improved,and the vehicle object detection framework WF-Center Net is obtained in this paper.Firstly,this paper studies the vehicle detection algorithm based on heat map.After the image features extracted from the backbone network are sampled upward through deconvolution,they are input to the three convolutional branch networks.Finally,the heap map describing the location of the object is obtained,and the size information and migration information of the object are regressed through the feature of the peak point.This end-to-end method not only solves the problem of high model complexity,but also improves the detection performance.At the same time,the frame regression from the peak point can eliminate the redundant non-maximum suppression process,which simplifies the calculation.Secondly,this paper chooses the improved DLA-34 as the backbone network.DLA-34 network can effectively aggregate spatial and semantic information of images.Although the basic convolution block as the residual structure can increase the depth of the network,the perception ability of the network to objects of different sizes varies greatly.Therefore,DLA-34 was improved in this paper to enhance the model’s perception ability to objects of different sizes by adding multiple convolution channels,thus improving the detection accuracy.In addition,the method of weighted feature fusion is also studied.In this paper,learnable weights are assigned to feature graphs of different scales so that different features have different influences on the output.In particular,the model in this paper adopts the feature fusion of FPN and TFPN,and finally adds them together.The weighted feature fusion method solves the problem that the different scale features have the same influence on the output in the traditional feature fusion method,and improves the feature utilization efficiency.Finally,this paper verifies the effectiveness of the algorithm.In this paper,the data related to six types of vehicles,namely cars,bicycles,buses,trains,trucks and motorcycles,were extracted from a large image dataset developed and maintained by Microsoft,and the experiments were carried out on Py Torch deep learning platform based on Ge Force RTX(?)3070.The experimental results show that compared with the mainstream methods,the proposed vehicle object detection algorithm based on heap map and weighted feature not only improves the detection precision,but also greatly simplifies the complexity of the algorithm while maintaining the real-time advantage.