| Forward vehicle recognition and distance detection is one of the key technologies for intelligent driving environment perception.In real scenarios,the variable vehicle types and scales,vehicle-vehicle occlusion,environmental occlusion and road pitch angle changes all increase the difficulty of forward vehicle recognition and distance detection to different degrees.The existing distance detection networks for specific targets have simple structures,serious loss of spatial information,and distance detection accuracy needs to be improved.Therefore,it is of great theoretical and practical significance to study forward vehicle recognition and distance detection based on deep learning.Based on the in-depth analysis of SSD(Single Shot Multi Box Detector)object detection and DORN(Deep Ordinal Regression Network)depth estimation algorithm,the paper improves the SSD algorithm for the problem of poor detection of small-scale vehicle targets and poor localization performance,and proposes a forward vehicle recognition and distance detection algorithm by multi-task learning with the improved vehicle detection algorithm and monocular depth estimation algorithm as two branches of the model.The main work and contributions of this paper are as follows.(1)To address the problem of poor detection of multi-scale vehicle targets by SSD,this paper proposes a multi-scale receptive field fusion module from two perspectives of receptive field and feature pyramid,and introduces an attention mechanism to reconstruct the feature pyramid to ensure that the features of each layer of the pyramid contain both feature representations of multi-scale receptive fields and semantic features encoding contextual information to improve the detection of small-scale targets.(2)To address the problem of poor localization performance of SSD,the cascade detection mechanism is introduced into the SSD algorithm to classify and regress Anchor twice to improve its localization accuracy.In addition,to gain the role of cascade detection,a feature alignment module with weights is designed to alleviate the drift problem between Anchor and features during cascade detection,and a weighted CIo U loss function is designed with reference to Anchor-free to effectively improve the regression accuracy.(3)For the problem that it is difficult to balance the correlation and discrepancy between the target detection task and the depth estimation task in the forward vehicle distance detection network,this paper introduces a multi-task attention network MTAN,which connects the target detection task and the depth estimation task in parallel,and proposes an end-to-end multi-task learning model for object detection and monocular depth estimation.In addition,to address the problem that the attention mechanism in MTAN is not sufficiently adaptive to the depth estimation task,a large kernel attention mechanism is introduced to improve MTAN,and a multi-task loss function adaptive weighting strategy is used to improve the accuracy of object detection and depth estimation.(4)For the problem that the depth values of non-vehicle regions in the 2D vehicle bounding box form interference to the distance detection,this paper proposes a distance measurement method based on K-Means optimization to effectively improve the accuracy of forward vehicle distance detection.In summary,a suitable algorithm for forward vehicle recognition and distance detection is proposed.This paper uses the forward vehicle recognition and distance detection dataset constructed by KITTI dataset for training and testing,and sets up several sets of comparison experiments,and the experimental results show that the algorithm in this paper has better advantages in accuracy and performs well in practical scenarios. |
