Deploying: Vehicle Deployment




Highway driving environment


I first began writing the post to share this video, but as I started writing, my thoughts expanded, making the post longer than expected.

As we can see in this video, the trained model is generalized to be able to run in a highway-like driving environment. Moreover, the training dataset covered environments with high speeds ranging from 80 to 120 km/h. However, CRN demonstrated effective performance across various scenarios, including comparably low speeds from 40 to 80 km/h and even urban driving scenarios!


Urban driving environment


To summarize the key aspects of this project:

Data Collection ➔ (Auto) Labeling ➔ Model Training ➔ Vehicle Deployment

I aimed to achieve these milestones with minimal effort and cost for the Proof of Concept (PoC). I just finished creating a barely working PoC and successfully confirmed that it is functioning! However, more effort is needed to improve this training cycle and the model.


Most importantly, I did not set Key Performance Indicator (KPI) to quantitatively measure performance.

Some parts of the network are not optimized for effective inference. Skip (residual) connections of ResNet might be better removed to lower the memory access cost. View transformation (voxel pooling) module can be further improved by using a fixed frustum grid during inference.

Lastly, TensorRT porting with INT8 quantization can be an effective way to significantly improve inference speed with minimal effort.