Goals

[x] Organize code and upload to github
[x] Fix the detection problem when yolo misidentifies car with truck for instance
[x] Fix orientation jumps

Github code upload

The code was uploaded to the oficial atlascar2 repository in a new branch where some changes of the master branch were executed and a new folder was added (atlascar2_perception). Inside this folder 3 new ros packages were created:

camera_perception
lidar_perception
risk_maps

GitHub - lardemua/atlascar2 at risk_maps

Detection problem

The solution involves creating a dictionary for every lidar label which contains the classes that had a higher IoU than the threshold as a key and their frequency as value. To test this solution the IoU threshold was lowered to 0,2. The following code spinet is the print of the dictionary where the first column is the lidar label. For this label 3 classes were associated with IoU > 0,2. The value next to each class is the number of times that the respective class was attributed. In the end the class with higher value of frequency is the one chosen.

{1053: {'car': 1}}
{1053: {'car': 2}}
{1053: {'car': 3, 'bicycle': 1}}
{1053: {'car': 4, 'bicycle': 1}}
{1053: {'car': 5, 'bicycle': 1}}
{1053: {'car': 6, 'bicycle': 1}}
{1053: {'car': 7, 'bicycle': 1}}
{1053: {'car': 8, 'bicycle': 1}}
{1053: {'car': 9, 'bicycle': 1}}
{1053: {'car': 10, 'bicycle': 1}}
{1053: {'car': 11, 'bicycle': 1}}
{1053: {'car': 12, 'bicycle': 1}}
{1053: {'car': 13, 'bicycle': 1}}
{1053: {'car': 14, 'bicycle': 1}}
{1053: {'car': 15, 'bicycle': 1}}
{1053: {'car': 16, 'bicycle': 1}}
{1053: {'car': 17, 'bicycle': 1}}
{1053: {'car': 18, 'bicycle': 1}}
{1053: {'car': 19, 'bicycle': 1}}
{1053: {'car': 20, 'bicycle': 1}}
{1053: {'car': 21, 'bicycle': 1}}
{1053: {'car': 22, 'bicycle': 1, 'truck': 1}}
{1053: {'car': 23, 'bicycle': 1, 'truck': 1}}
{1053: {'car': 24, 'bicycle': 1, 'truck': 1}}
{1053: {'car': 25, 'bicycle': 1, 'truck': 1}}
{1053: {'car': 26, 'bicycle': 1, 'truck': 1}}
{1053: {'car': 27, 'bicycle': 1, 'truck': 1}}

Orientation jumps

The orientation is calculated as stated previously based on the current and previous position of the object. This orientation is then smoothed by using an average filter which computes the average orientation in the past 5 estimations. If the object stops after the movement the orientation will be the previous one. If the object has no movement the orientation will remain stochastic (eg.: the pedestrian in the video).

orientation.webm