traffic_motion_predictor

Given 1 second of traffic and an agent's starting position, predict its trajectory 8 seconds in the future

The pitch

---

Findings

name	Pretrained CNN	Backbone CNN	Type	Learning Rate	Momentum	minFDE_val	minADE_val	Epochs	Batch Size	Mixed Precision
simple_cnn_1	Yes	mobile-net	SimpleCNN	0.0001	0.9	228	1622	160	16	N
simple_cnn_1	Yes	mobile-net	SimpleCNN	0.0001	0.9	266	2158	92	16	N
simple_cnn_2	Yes	mobile-net	SimpleCNN	0.001	0.9	399	3032	92	16	N
simple_cnn_3	Yes	mobile-net	SimpleCNN	0.0001	0.9			92	28	Y

name to id:

• simple_cnn_1: simple_cnn_rofjQV2NUK1trG6vmlqOSEDd7bHr9OxUHzAVEjV7d
• simple_cnn_2: simple_cnn_rVW9hZ6wZcpHqLIen3Tio1r8haPBGTmtcUitSfVfJ
• simple_cnn_3: simple_cnn_rSFaNYXVMP9P8hPScrKBmgGDA0oZSgELrCYk6Iqod

Setup

Setting up Python

1. install python 3.10
2. (optional) setup virtual environment

linux/mac:

python -m venv .env
source .env/bin/activate

windows:

python -m venv .env
./.env/Scripts/activate

3. install dependencies

python -m pip install requirements.txt

Setting up NuScenes mini dataset

Download the Mini Dataset

mkdir -p data/sets/nuscenes  # Make the directory to store the nuScenes dataset in.
wget https://www.nuscenes.org/data/v1.0-mini.tgz  # Download the nuScenes mini split.
tar -xf v1.0-mini.tgz -C data/sets/nuscenes  # Uncompress the nuScenes mini split.

(for windows the easiest way is to just run these commands in wsl)

Adding the US Map expansion pack

1. Open the downloads page and go to Full Dataset (v1.0) > Mini and click US
2. download US Map expansion pack (v1.3)
3. decompress both packages and drop basemap/, expansion/ and prediction/ (all from the maps expansion pack) into v1.0-mini/maps/

Jupyter Notebook