Unstructured Environment Detection And Recognition Based On Double Sensor Style

The research of unmanned driving on structured roads has been gradually matured and applied.However,there are still many technical bottlenecks in the perception of the environment on unstructured roads.One of the major significant challenges is the depiction of unstructured roads as passable areas and detecting negative road obstructions.As sensors commonly used in unmanned driving research,LIDAR and vision cameras have their advantages for different scenarios.Considering the application scenarios and the influence of practical factors,as sensors commonly used in driverless research,LIDAR and vision cameras both have their advantages for different scenarios.They were considering the application scenario of this research and the influence of practical factors,this paper proposes passable area depiction based on 16-line LIDAR and vision camera-based negative obstacle detection.The primary research of the thesis includes:(1)For unstructured roads with complex and variable conditions,the effectiveness of commonly used structured road detection methods is significantly reduced when applied to unstructured roads.This paper,proposes a KD-Tree edge point extraction method that uses maximum variance as the dimensional division and median as the division value.Considering the advantages and disadvantages of the least squares method and the random sampling consistent algorithm,combining the random sampling consistent algorithm and the least squares method is proposed to achieve the road edge line fitting.The experimental results show that the method can complete the detection of unstructured road passable areas more accurately.(2)In response to the problem of sparse point clouds or no reflection data when LIDAR scans potholes on the road,this paper proposes to use vision camera sensors to detect negative obstacles on the road.In this paper,we use a vision camera to capture image samples of real roads and mark the locations of negative obstacles on the road to complete the construction of the negative obstacle dataset.The dataset is input to the YOLOv5 network to train the negative obstacle detection model,and the precision of the model is 90.6%,the recall(Recall,R)is 89.6%,and the average precision means m AP＿0.5 and m AP＿0.5:0.95 are 91% and53.2%,respectively.The experimental data shows that the model obtained from YOLOv5 training can meet negative obstacle detection requirements.(3)Considering the influence of image noise and background factors during practical applications,this paper proposes a new ECA-YOLOv5 model,which embeds the ECA attention mechanism in the backbone feature extraction layer and the neck feature fusion layer of the YOLOv5 algorithm.This method uses limited attention resources to filter out essential features from a large dataset quickly.It ignores the influence of background factors,making ECA-YOLOv5 pay more attention to critical features when learning target features to improve the performance of network training.The negative obstacle detection model obtained from the ECA-YOLOv5 network training has a precision of 91.2%,a recall of 93.1%,and average accuracy means of m AP＿0.5 and m AP＿0.5:0.95 of 93.1% and 58.6%.Compared with the unimproved YOLOv5 model in P,R,m AP＿0.5,and m AP＿0.5:0.95 are improved by 0.6%,3.5%,2.1%,and 5.4%.