Merge pull request #43 from una-auxme/4-feature-pylot-planning-PAF21/…

…2-Perception

4 feature pylot planning paf21/2 perception

doc/03_research/research_PAF_23/Research_Pylot-Planning_PAF21-Perception.md

@@ -0,0 +1,67 @@
+# Sprint 0: Research Samuel Kühnel
+
+## Pylot
+
+## Planning
+
+- 4 different options
+- **Waypoint Planner**: Auto follows predefined waypoints. It recognizes traffic lights and stops at obstacles, but cannot avoid them
+- **Freenet-Optimal-Trajecotry-Planner**: CPP code with Python wrapper ([GitHub](https://github.com/erdos-project/frenet_optimal_trajectory_planner))
+→ Predefined line that is used for orientation → Can avoid obstacles!
+
+![freenet_gif](https://github.com/erdos-project/frenet_optimal_trajectory_planner/raw/master/img/fot2.gif)
+
+- **RRT\*-Planner**: RRT* algorithm for path planning ([GitHub](https://github.com/erdos-project/rrt_star_planner))
+  - Creates random nodes
+  - Adds nodes to the graph that are not blocked by objects on the road
+  - Generally terminates as soon as a node is found in the target area
+  - RRT*: Searches for the shortest path
+
+![rrt_star_gif](https://github.com/erdos-project/rrt_star_planner/raw/master/img/rrtstar.gif)
+
+- **Hybrid A\* planner**: Hybrid A* algorithm for path planning ([GitHub](https://github.com/erdos-project/hybrid_astar_planner))
+  - Calculates the shortest path between two nodes from a graph
+  - Similar to Dijkstra's algorithm
+  - Nodes are estimated based on their costs and promising nodes are selected first
+  - Hybrid A* algorithm: Not always optimal solution, but in the neighborhood of the optimal solution.
+
+![hybrid_astar_gif](https://github.com/erdos-project/hybrid_astar_planner/raw/master/img/straight_obstacle.gif)
+
+## PAF 21-2
+
+### Perception
+
+### Obstacle detection
+
+- Detect objects via semantic lidar sensor
+- Provides x and y coordinates, as well as distance value
+- Additional information on position change in a time interval → Calculation of speed possible
+- Detects the object and returns either the value "Vehicle" or "Pedestrian"
+
+### TrafficLightDetection
+
+![diagramm.png](https://github.com/ll7/paf21-2/raw/main/docs/imgs/trafficlightdetection_diagram.jpg)
+
+- **FusionCamera** saves images from **RGBCamera** and **DepthCamera** with timestamp and then synchronizes with **SegmentationCamera**
+- Neural network based on [ResNet18](https://pytorch.org/hub/pytorch_vision_resnet/) (predefined PyTorch network)
+- Generally only traffic lights up to 100m distance
+- Canny algorithm to filter contours
+
+### Problems and solutions
+
+- Red background distorts traffic light phase detection → **Solution**: Narrow section of the traffic light image for phase detection
+- Yellow painted traffic lights distort traffic light phase detection → **Solution**: Filter out red and green sections beforehand using masks and convert remaining image to grayscale and add masks again.
+- **Problem without solution**: European traffic lights can sometimes not be recognized at the stop line.
+
+## Resumee
+
+### Perception
+
+- Status quo: LIDAR Sensor and Big Neural Network
+- Possible focus: Using smaller (pretrained) models to improve overall performance fast
+- Taking known issues into account
+
+### Planning
+
+- Currently decision tree for evaluating the current position
+- Trying out different heuristics → already given as repo