Research On Algorithms Of Lane Detection And Pedestrian Vehicle Detection Based On Monocular Camera

With the rapid development of our country’s economy,automobiles have entered the mass family and become the main means of transportation.At the same time,with frequent traffic accidents,people put forward higher requirements for the safety of vehicle driving.The advanced driving assistance system senses the surrounding environment of the vehicle,and warns the driver in real time in case of dangers,which reduces the occurrence of accidents.The key algorithms of vision-based advanced driving assistance system are the lane detection and pedestrian vehicle detection algorithms based on monocular camera.Therefore,it is of great significance to study the lane detection and pedestrian vehicle detection algorithms based on monocular camera to improve the performance of the algorithms.At present,lane detection and pedestrian vehicle detection in driving scene are challenging due to the complexity of the scene,and the need of balance between robustness and complexity.A real-time lane detection algorithm and an improved algorithm to overcome the problem of low recall rate of pedestrian and vehicle detection are proposed in the thesis.The main work of the thesis includes the following.1.A lightweight lane detection method is proposed in the thesis.Line segments are detected by EDLines features.The vanishing point is estimated with the valid line segments.Then the valid line segments are connected to form smooth long lane edge lines.The left and right lane of the current lane are detected by the joint probability of the lane boundary pairs.2.YOLOX is used as baseline framework for pedestrian and vehicle detection.A spatially separable self-attention module is introduced as a global attention module in the top most feature map to learn rich global context information.And a multi-scale spatial attention module is proposed to learn rich local context information,which guides the model to focus on the target area and improves the recall rate of the algorithm.3.In view of the lack of global information in adaptively spatial feature fusion,a global adaptively spatial feature fusion GASFF is proposed,which improves adaptive spatial feature fusion through a lightweight global attention mechanism.4.Based on the proposed YOLOX-GASFF model and LSTR model,a multi-task network for detecting lane and pedestrian and vehicle is designed.