PROJECT 04 · Computer vision · Robotics simulation

Autonomous Navigation in CARLA
Perceive. Plan. Act.

A simulated autonomous-driving agent that combines vision, lane detection, and control. I benchmarked two object-detection approaches and built a route-following controller around the simulation environment.

CARLA / PERCEIVE → PLAN → ACT

PERCEIVE
↓ PLAN ↓
ACT

Bhanu Teja Malineni04 / 04

RoleProject developer

PeriodProject work

ContextCARLA simulation

FocusComputer vision

The idea

Autonomous navigation requires a system to interpret a scene, choose what to do, and control movement. I built a Perceive–Plan–Act agent in CARLA simulation and compared detector speed as part of choosing a real-time perception strategy.

How it works

I benchmarked SSD at 25–30 FPS against Faster R-CNN at 3–6 FPS, with confidence above 0.95 reported for Faster R-CNN. The navigation pipeline used OpenCV for lane detection, a PID controller for steering behavior, and a finite-state machine to manage driving states.

01PerceiveCompare detector speed and use OpenCV lane detection.

02PlanUse a finite-state machine to select the next driving behavior.

03ActApply PID control to follow the planned route in simulation.

Reported outcomes

25–30 FPSReported SSD processing rate.

3–6 FPSReported Faster R-CNN processing rate; confidence >0.95.

98–100%Reported route completion across 100-waypoint test runs.

SimulationTested in CARLA; results describe simulated runs.

Performance values are reported in the source resume and refer to simulation/testing conditions. They do not describe a road-tested vehicle.

Tools & methods

PythonCARLAOpenCVSSDFaster R-CNNPID controllerFinite-state machine

More about the project

The detector benchmark informed the real-time perception choice, while lane detection, the controller, and state machine formed the rest of the navigation loop. The project was tested through waypoint-based runs in CARLA simulation.

What I took from it

The project brought perception and control together. Benchmarking made the speed-versus-detection tradeoff visible, while the controller and state machine turned visual cues into route-following behavior.