Research On Object Detection Algorithm For Autonomous Driving Scene Based On Improved YOLOv4 Algorithm

In recent years,with the increasing number of cars,the traffic problems faced by the traditional automotive industry have become more prominent.Therefore,autonomous driving technology has received increasing attention and research from governments and businesses as an important solution.Accurate detection of vehicles and pedestrian targets in the traffic environment is necessary before achieving autonomous driving tasks.Although vision-based object detection algorithms have made rapid advancements,in complex traffic scenarios,dense occlusions among targets and a large number of small targets lead to significant instances of missed detections,making it challenging to meet the required detection accuracy.Additionally,the size of object detection models is often too large to meet practical deployment requirements.To address the aforementioned issues,this study focuses on deep learning-based object detection techniques and conducts research on object detection algorithms specifically designed for autonomous driving scenarios.The main research contributions are as follows:To address the problem of low detection accuracy and severe missed detections in the YOLOv4 algorithm,this paper first embeds the SimAM into the residual modules of the backbone network.This enhances the feature extraction capability of the backbone network,allowing the model to pay more attention to critical feature details in the images.Furthermore,the original Mish activation function in the residual module is replaced with the ACON-C activation function,which provides better non-linear expression capability,enabling the residual module to activate adaptively and thereby improving the network’s performance.Finally,transfer learning technology is utilized by training the model with a pre-trained model as the initial state,and the SimAM-YOLOv4 algorithm model is obtained.The experimental results on the KITTI dataset show that SimAM-YOLOv4 can obtain 91.19%mAP and 32 FPS.To address the challenge of deploying large detection models,the study proposes a pruned model called Slim-SimAM-YOLOv4.Firstly,sparse training is applied to the SimAM-YOLOv4 algorithm model,enhancing the sparsity of convolutional channels and causing many scaling factors in the model to approach zero.This differentiation between important and unimportant channels prepares the model for pruning.Then,channel pruning is performed on the sparse SimAM-YOLOv4 model by setting different global channel pruning thresholds,effectively removing redundant channels while preserving the useful ones to the maximum extent,thereby reducing the model’s complexity and parameters.Finally,the model after channel pruning was layer pruned,and the network layers with lower mean scaling factors were pruned to further reduce the complexity and model parameters of the network model.The pruned network model achieves significant optimization in terms of both width and depth.The pruned network model undergoes retraining and fine-tuning to obtain the final Slim-SimAM-YOLOv4 object detection algorithm model.By simultaneously improving accuracy and speed,the size of the network model is greatly compressed,reducing the hardware load.Experimental results on the KITTI dataset demonstrate that the SlimSimAM-YOLOv4 model achieves a maximum mAP of 89.56% and a frame rate of 54 FPS,while compressing the model size to 69.2MB,reducing it by 71.6%.