top of page

MAZE  SOLVING  ROBOT

The Project

The project's aim was to find a goal location within a maze using only monocular Camera input. There were 2 phases of the project. 

Exploration Phase: This phase allowed us to explore the maze without being timed. 

Navigation Phase: A timed phase where we had to minimize the time of reaching the goal location

​

Evaluation Criteria

  • Time taken to reach the goal position

  • Accuracy of the goal position

  • Method implementation

  • Submission time

Challenge

  • The goal position was defined by 4 Camera views (Left, Right, Front, Back) taken from the goal position 

  • The 4 camera views were revealed after the Exploration phase was over

  • The accuracy of the position reaches was measured by the squares defined around the goal location which was not visible to the participants of the challenge

Goal Position Constraints

Method : Exploration Phase 

  • First, within see() Function, we check if we are within the exploration phase

  • If so, we write each frame to self.all_fpv and write every 5th frame to the explore_phase folder

  • Subsequently, we perform template matching on consecutive frames by calling our self.slam attribute which is instantiated with our SLAM() object from VisualSlam.py

  • If enough matches are found, we call the get_pose() method from the self.slam object.

  • Then, we check via our last action if a rotation is performed, i.e, Action.LEFT or Action.RIGHT or if a translation is performed- i.e, Action.FORWARD or Action.BACKWARD

  • If only a rotation is formed we extract and use only the rotation matrix. If a translation is performed we use the full transformation matrix.

  • Else, if we were idle, i.e, Action.IDLE, nothing is done.

  • We store the current pose in self.cur_pose

  • This is then appended to the self.estimated_path list.

  • Furthermore, we append the frame count and the calculated pose to the self.location_db dictionary.

  • self.estimated_path is used for the pose graph and self.location_db is used for place recognition

  • The process hierarchy is shown in the flowchart

  • We use ORB and FLANN for feature extraction and matching ( These are used for their speed and efficiency when compared to sift or surf for extraction and Bfmatcher for matching )

Architecture of explore phase

Method : Navigation Phase 

  • Pass annotated frame to self.fpv to be displayed within pygame

  • Check if the graph is generated,

    • If not, create Bag of words object, self.bovw​

    • Make the codebook via sel.bovw.makecodebook()

    • Perform tf_idf calculations via self.bovw.generate_tf_idf_vectors()

    • Compare target images to other images via cosine similarity via self.bovw.search()

    • The search function will return an integer with the frame count from the explore_phase database, this
      frame is the closest match.

    • Parse this integer to the location_db {} dictionary as a key and store the returned value

    • The list of 4 returned values is passed to visualize_paths() to generate the graph

    • The graph highlights these 4 points in red and computes the centroid in blue, this should correspond to our target location


       

  • The video feed shows some matches and the target location to demonstrate how this annotation helps us.

Architecture of navigation phase
Graph drawn from Visual Odometry

Underlying Math

Epipolar Geometry:

bottom of page