Application of Improved YOLOV4 in Intelligent Driving Scenarios

Abstract

With the development of unmanned technology, the technical innovation of invehicle vision detection system is also getting faster and faster, while the improvement of algorithm accuracy often brings an increase in the number of parameters and poor real-time performance. In his paper, the optimization of the algorithm structure of YoLoV4 target detection is achieved by using MobileNet-v3 instead of the CspDarkNet53 master Network, which has the inverse residual structure of linear bottleneck, while the lightweight attention mechanism is added to the feature extraction process, and the learning degree of feature channels is enhanced; due to the long computation time of sigmoid, it also uses ReLU6(x+3)/6 is used to approximate the original activation function due to the long computation time of sigmoid; the system parameters are reduced by constructing a depth-separable convolution instead of the normal convolution in PaNet. Meanwhile, this paper improves the original upsampling method by using dual cubic interpolation, which makes the image more smooth, less image loss and more accurate feature extraction during he upsampling method. The map% is improved from 79.1% to 81.2% on the voc dataset, reaching 58.14 FPS.