Abstract:

Visual simultaneous localization and mapping(SLAM) plays an important role in autonomous driving. However, it is difficult for the current visual SLAM algorithm to solve those complex traffic scenes where few nearby textures but a large amount of dynamic object occlusion such as moving cars exists. Therefore, this paper proposes a visual SLAM algorithm based on key objects. The algorithm consists of five parts—key object extraction, system initialization, tracking based on key targets, pose solution and local bundle adjustment(BA).

First of all, based on stationary targets detected by the environment perception algorithm in typical traffic scenes such as traffic signals and signs, this paper performs feature extraction in stationary targets and uses the quad tree to homogenize the feature points. According to the number of feature points, the key targets are screened and the process of key object extraction is finished.

Secondly, the key target matching between the two continuous frames is performed in order to help system initialization. The matching criteria consist of the center, the height and the width of the key targets. System initialization is launched when the number of the matched key targets satisfies the need. RANSAC algorithm is used in feature point selection to compute the initial pose. A local map and map points are established at the end of the system initialization.

Then, the constant motion model is applied to predict the center point of the key targets in the following frame. Besides the height and the width of the key targets, the number of the matched points is also taken into account to decide which key target in the current frame matches the corresponding one best.

Furthermore, after data association is completed, the camera pose is estimated based on the feature points and the center point of each key target. Reprojection error is chosen to optimize the pose and distinguish the inliers and outliers in the process of optimization iteration.

Lastly, local BA is utilized in local mapping thread in order to reduce accumulative error. Levenberg-marquardt(LM) algorithm is used as the optimization algorithm and the optimization variables include the pose of the key frame, the center point of each key target and the 3D coordination of feature points in key targets.

To verify the performance of the proposed algorithm, the typical scenes in ApolloScape dataset are chosen. Experiments show that the proposed algorithm can effectively solve the problem of localization failure in some near-texture missing environments, and maintain a high localization accuracy, which has a good environmental adaptability.