Merge pull request #42 from una-auxme/9-paf_21_1_perception

9 paf 21 1 perception

doc/06_perception/09_paf_21_1_perception.md

@@ -0,0 +1,38 @@
+# Paf_21_1 - Perception
+
+## 1. Architecture
+
+![image](https://github.com/una-auxme/paf23/assets/102369315/07328c78-83d7-425c-802e-8cc49430e6c1)
+
+### **Key Features**
+
+- Object-Detection
+- TrafficLight-Detection
+
+## 2. Sensors
+
+- Position
+- Velocity
+- RGB-Image
+- Depth-Image
+- Semantic-Image
+
+## 3. Object-Detection
+
+- Objects are detected using semantic and depth images
+- First they use masks and filters on the semantic image to identify pedestrians and vehicles
+- Then they use the depth image and the position and velocity data to calculate the relative positions of each object
+
+## 4. TrafficLight-Detection
+
+- They use the semantic image and apply masks and filters to find traffic lights
+- Then they use this information anlognside with the depth image to cut out a specific patch of the image
+- After that they cut the same patch from the RGB Image (32x32x3) which is than classified by a simple Convolutional Neural Network (CNN)
+- They use the classes Backside, Green, Yellow and Red
+
+## 5. Conclusion
+
+- Some parts of their systems, like the CNN, are probably usefull to us
+- Since their entire perception is based on semantic and depth images it will be vewry hard for us to adopt their concepts
+- We would have to build our own semantic and depth cameras which is a very complex task
+- All in all the concepts of paf_21_1 regarding the perception aren´t to usefull for our project