Ab­stract

For ro­bust ob­ject de­tec­tion on LiDAR data, neural net­works have to be trained on di­verse datasets that con­tain many dif­fer­ent en­vi­ron­men­tal in­flu­ences like rain, snow, or fog. To this date, few datasets, with those fea­tures, are avail­able while there exist many datasets recorded under per­fect weather con­di­tions. Re­pur­pos­ing those datasets by sim­u­lat­ing ad­verse en­vi­ron­men­tal con­di­tions on top of them and train­ing net­works with the re­sult­ing en­hanced datasets, is in­tended to lead to more ro­bust neural net­works. In the fol­low­ing we pro­pose mod­els to re­al­is­ti­cally sim­u­late the ef­fects of rain, snow, and fog on LiDAR datasets based on phys­i­cal and em­pir­i­cal fun­da­men­tals. Then we pa­ra­me­ter­ize our sim­u­la­tion to best fit real LiDAR data that was cap­tured in those en­vi­ron­ments, in order to achieve a highly ac­cu­rate sim­u­la­tion. Fi­nally, the im­pact of ad­verse weather on neural net­work de­tec­tion per­for­mance is demon­strated.