Future_work.md

Future work

This project lays a solid foundation for future improvements. We will only address the Gradient merge & Centroid detection method.

Utilize the planar LIDAR

By smartly clustering the points and estimating centroids in the planar LIDAR scan, we can get an additional, more precise estimate of the car distance, as the camera LIDAR sometimes returns 0 in the centroid. Furthermore by detecting and filtering the track boundaries can mark the regions of interest as only points from the car and obstacles will be left in the scan. By filtering out the static objects within the frames, the presence of the vehicle in the scene can be detected (assuming the vehicle is moving).

Bounding box estimation

By combining the data from all sensors and prior knowledge of the car ratio (rectngular) a bounding box can be estimated. This can not only help to highlight the detection but can be used for further processing.

Use convolutional neural network

A light CNN model can be trained to estimate the vehicle heading. The input to the model should be the cropped image obtained by the estimated bounding box.