Free Space Detection Method Based On SENet Attention

The free space based on deep learning object detection still has some uncertainties when detecting static boundary or dynamic obstacle boundary, this will make it impossible to make effective track planning and state decision for vehicle running state, therefore, we need to modify the semantic edge information by matching the results of curb, lane line and target box and define the free space from the vector envelope or raster diagram. Secondly, the detection of driving area may be affected by light, dust, heavy snow and fog, therefore, it is necessary to fully combine vision and radar for obstacle detection to ensure its detection stability. In the process of free space detection, sensors such as cameras and lidar are easily affected by the environment, light conditions and extreme weather will affect the accuracy of the detection results of free space, in the traffic jam section, the occlusion caused by the surrounding vehicles will also affect the accuracy of the detection results of the free space. This paper refers to a CNN data fusion architecture, called RoadSeg, it can from the RGB image feature extraction and fusion, and deduce the surface normal information supplied, so as to realize the free space of accurate detection [1]. We introduced SENet on RoadSeg[2] called SE-RoadSeg to make some results more accurate.