Milo Henry Lovelace Knowles

abstract:Many autonomous robots, such as self-driving cars, must navigate through and interact with dynamic environments. Commonly used approaches in SLAM and visual odometry assume a static world, which can lead to inaccurate scene reconstructions and pose estimates if many dynamic objects are present. This project aims to mitigate these issues by tracking and modelling the motion of dynamic objects in a scene. By incorporating an object detector into our SLAM pipeline, we can detect dynamic objects in a scene and mask them during feature tracking. In addition, we can fuse temporal measurements of each object over time to infer its motion. These motion estimates can then be used for predictive planning and obstacle avoidance.