Research On Obstacle Detection Algorithm For Autonomous Driving

The gradual maturity of sensor technology and 5G communication technology provides a good development environment for autonomous driving technology.Autonomous driving technology can assist drivers in making decisions and improve the safety and convenience of vehicles.With the development of deep learning theory,the most critical obstacle object detection technology in the field of autonomous driving has also become one of the current research hotspots.The automatic driving obstacle object detection technology can timely and accurately detect the obstacle target in the driving area of the vehicle,which improves the driving safety of the vehicle,and has very important research significance.However,the obstacle targets in the autonomous driving scene are usually small targets,and the memory of the on-board computer is relatively limited.The existing object detection algorithms have low detection accuracy for small targets and a large amount of computation,which is difficult to meet the needs of automatic driving obstacle object detection tasks.Therefore,this paper mainly focuses on the two core problems of low detection accuracy of small targets and limited memory of the on-board computer,and proposes a real-time object detection algorithm based on multi-scale feature fusion,which realized high-precision real-time detection of obstacle targets in autonomous driving.The research contents and innovations of this paper are as follows:(1)An anchor matching strategy based on sample supplementation is proposed.In the singlestage object detection algorithm,the anchor matching strategy is very important.The current anchor matching strategy only matches one anchor as a positive sample for each target,and does not consider the matching degree of the small target and the anchor.The poor problem reduces the learning effect of the network on small object features.This paper proposes an anchor matching strategy based on sample supplementation.This strategy will supplement a small amount of positive samples for objects with a general matching degree,and supplement more positive samples for objects with poor matching degree.The number of positive samples improves the learning effect of the network on small object features.Compared with the default anchor matching strategy,the anchor matching strategy proposed in this paper improves the full object detection accuracy by 3.7% and the small object detection accuracy by 12.1%,which greatly improves the detection effect of small objects.(2)A lightweight object detection network based on balanced feature pyramid is proposed.The current object detection network uses the feature maps of multiple downsampling for object detection,while the targets in the automatic driving obstacle object detection task are mostly small targets,and the small targets are prone to feature loss after multiple downsampling.The features of the small objects in the figure are small,resulting in a low detection accuracy of the small objects by the network.At the same time,the existing object detection network has a large amount of parameters and computations,and it is difficult to complete the real-time detection of obstacle targets,and it is necessary to reduce the weight of the network structure.In order to improve the detection accuracy of small objects and improve the network’s ability to represent small objects,this paper designs a balanced feature pyramid network(Balanced Feature Pyramid Networks,BFPN)and an attention-based balanced feature pyramid network(Squeeze-and-Excitation Balanced Feature Network).Pyramid Networks,SE-BFPN)two feature fusion structures.BFPN uses five feature maps of different scales for feature fusion,which supplements the extremely shallow target shape features and target location information that are very important for small object detection in the feature maps.SE-BFPN introduces an attention mechanism on the basis of BFPN,which can enhance the important features in the feature map and weaken the useless features,and improve the efficiency of feature fusion.Compared with the FPN feature fusion structure,BFPN improves the full object detection accuracy by 4.3% and the small object detection accuracy by 6.2%,and SE-BFPN improves the full object detection accuracy by 5.0% and the small object detection accuracy by 6.6%.On the other hand,in order to solve the problem of limited memory in the on-board computer,this paper uses a depthwise separable convolution.By controlling the number of channels and the resolution of the feature map during the convolution operation,the detection accuracy of the algorithm is not significantly reduced.It reduces the network parameters by more than 50% and the computation volume by more than 60%.The lightweight object detection network based on the balanced feature pyramid proposed in this paper can balance the target location information and target semantic information in the feature maps of each scale,and effectively improve the detection accuracy of the algorithm for small targets.The amount is small,which greatly improves the detection speed of the algorithm.(3)A non-maximum suppression algorithm based on position weighted correction is proposed.In the task of automatic driving obstacle detection,the position of the small object is usually not accurate enough.The general non-maximum suppression algorithm only retains the detection with the highest confidence,and cannot correct the position of the small target.The non-maximum suppression algorithm proposed in this paper can use the position information of redundant detection frames to perform weighted position correction on the results,which further improves the positioning accuracy of small targets.Compared with the general non-maximum suppression algorithm,the non-maximum suppression algorithm based on position weighted correction improves the detection accuracy of small objects by2.3%,and reduces the false detection rate of the algorithm for small objects.