Obstacle Avoidance System Design and Path Planning for An Unmanned Vehicle

1. Developed a path-planning system based on Rapidly-Exploring Random Tree (RRT)
2. Integrated the path-planning system with ultrasonic system
3. Developed a specialized auto-exposure method for stereo camera
4. Developed a method to modeling obstacles detected by stereo camera

Path-planning System

I developed the path-planning system based on Rapidly-Exploring Random Tree (RRT). The path-planning system takes the coordinates of the obstacles detected by stereo camera, creating a path to avoid the obstacles.

Here is an example of my RRT implementation:

step 0: tree reaches goal	step 1: intial path
step 2: shorten path	step 3: trim unnecessary points

The source code of RRT and a simple GUI(developed using Qt) are on Github

In real situation, the path-planning system only discovered the square area around the vehicle, and the area followed the position of the vehicle. In addition to the path points generated by RRT, the path-planning system utilized hermite curve interpolation to generate full path. See Case 1 in the result page for real test.

^TOP

Integration of Path-planning And Ultrasonic Systems

I integrated the path-planning system with the ultrasonic sensor system, which was employed as a real-time obstacle avoidance method. The characteristics of these systems are:

RRT-based Path-planning System:

Pro:

• It took the obstacles detected by stereo camera, so the range of detection was much larger than ultrasonic sensors
• It handled multiple obstacles in systematic way

Con:

• It cost large amounts of computing resources
• The generated path might differ largely from the previous path due to its intrinsic randomness

Ultrasonic Sensor System:

Pro:

• It took the obstacles detected by ultrasonic sensors and utilized fuzzy rules to avoid obstacle, so it runs faster.

Con:

• It can only avoid obstacles within 1 meter.
• Its scope of obstacle avoidance was localized

I developed a decision flow to integrate these systems so that they complement each other:

See Case 2 in the Result page for real test.

^TOP

Auto-exposure Method

Since the vehicle walks on the ground, the lower part of stereo camera should be ground where the exposure value is close to the obstacle. The specialized auto-exposure method can be expressed as following steps:

1. Weighted Average Brightness = (Average greyscale of the lower 1/3 part of the image) * 0.7 + Averge greyscale of other part * 0.3
2. Iterate changing the exposure setting of stereo camera so that this weighted average brightness falls between 96~160(8-bit greyscale).

Built-in auto-exposure method	Specialized auto-exposure method

With the specialized auto-exposure method, the images captured by stereo camera contained more details. Consequently, the result of obstacle detection was improved. This algorithm was inspired by a common technique of photography which was my personal interest.

^TOP

Modeling Obstacle Detected By Stereo Camera

Position vectors of all pixels in the left eye image were obtained by the disparity map calculated by SRI Stereo Engine accompanied with our stereo camera. The disparity map is illustrated as below:

Left Eye Image	Disparity Map by SRI Stereo Engine

All pixels of the disparity map were put to a grid in front of the vehicle according to their coordinate, like:

If the cumulative pixels of a block were higher than a specific threshold and the average height was higher than ground, that block was considered as an obstacle, represented by its central point, and sent to main station program for path planning.

^TOP