Research On Road Target Detection Algorithm Based On Deep Learning

Under the "new four" and "carbon neutral,carbon peak" dual carbon goals,autonomous driving as one of the core of "intelligence" has been developed rapidly.Road target detection(vehicle detection,pedestrian detection,etc.)is a relatively basic and important technology in the task of autonomous driving perception,however,the scenario of autonomous driving is characterized by a complex environment and changing conditions,and target occlusion and lighting effects can lead to the failure of common detection algorithms.It can be seen that the road target detection algorithm still has a large room for improvement.In this paper,we do the following work on how to control the false detection of targets and improve the detection accuracy of targets.(1)To enhance the control of vehicle misdetection by the autonomous driving platform,an enhanced negative-sample vehicle detection algorithm is proposed.First,to coordinate the consistency of classification and regression,the loss function is improved based on Generalized Focal Loss to optimize the two branches;second,an adaptive training sample selection strategy is introduced to better balance positive and negative samples.Finally,a negative sample extraction and fusion module is designed to make full use of high-quality negative samples,and a semi-supervised learning method with optimized false detection is used to iteratively train the network model.The experimental results on KITTI and UA-DETRAC datasets show that the algorithm has significantly improved accuracy and false detection control compared with other mainstream target detection algorithms.(2)In order to improve the detection accuracy of the algorithm in complex scenes,a road target detection algorithm based on deformable convolutional network is proposed by improving the network structure.First,the backbone network Res Net50 is modified by deformable convolution to improve the detection capability of the model for target deformation;second,the attention module Global Context block is added to optimize the modeling capability of the global context;meanwhile,the expression capability of the model detection head is improved by unifying the multiple attention mechanisms.After that,Soft-NMS(Soft Non Maximum Suppression)algorithm is introduced for bounding box fusion to solve the occlusion problem.Finally,the experimental results on Huawei SODA10 M dataset show that the algorithm can effectively improve the detection accuracy compared with other mainstream target detection algorithms.(3)A home-made road target dataset with school neighbourhood and school interior as the main scenarios is developed,and the algorithms proposed in this paper are trained and tested on this dataset.The experimental results show that the two algorithms proposed in this paper have equally good detection performance in real complex scenarios.