Create pylot docs

doc/03_research/02_perception/04_pylot.md

@@ -0,0 +1,54 @@
+# Pylot - Perception
+
+**Authors:** Maximilian Jannack
+
+**Date:** 2.11.2023
+
+---
+
+## [Detection](https://pylot.readthedocs.io/en/latest/perception.detection.html)
+
+### Obstacle detection
+
+Pylot provides two options for obstacle detection:
+
+1. Obstacle detection operator that can use any model that adheres to the Tensorflow `object detection model zoo`
+   - By default, three models that were trained on 1080p CARLA images (`faster-rcnn`, `ssd-mobilenet-fpn-640`, and `ssdlit-mobilenet-v2`) are provided
+   - Models that have been trained on other data sets can be easily plugged in
+2. Operator that can infer any of the EfficientDet models (not trained on CARLA data, but on the COCO dataset)
+
+### Traffic light detection
+
+Uses `Faster RCNN weight` (trained on 1080p CARLA images)
+
+### Lane detection
+
+Uses the `Lanenet` model ([repo](https://github.com/MaybeShewill-CV/lanenet-lane-detection)) or canny edge detector.
+
+---
+
+## [Obstacle Tracking](https://pylot.readthedocs.io/en/latest/perception.tracking.html)
+
+For tracking obstacles across frames.
+Uses the `DaSiamRPN` model ([repo](https://github.com/foolwood/DaSiamRPN)) to serially track multiple obstacles.
+
+---
+
+## [Depth Estimation](https://pylot.readthedocs.io/en/latest/perception.depth_estimation.html)
+
+Uses stereo cameras to estimate depth with the `AnyNet` model ([repo](https://github.com/mileyan/AnyNet)).
+Configurable camera distance between left and right.
+
+---
+
+## [Segmentation](https://pylot.readthedocs.io/en/latest/perception.segmentation.html)
+
+Different approaches, such as using the `DRN` model ([repo](https://github.com/ICGog/drn)) for segmenting camera images.
+No model with training on CARLA data available, output of segmentation component not used in Pylot right now.
+
+---
+
+## [Lidar](https://github.com/erdos-project/pylot/blob/master/pylot/perception/point_cloud.py)
+
+Pylot contains a few helpful function for handling point clouds from the Lidar sensor.
+It can for example merge two point clouds or map the points to a camera image.