Table 4 LSTM approach
From: Object detection for automotive radar point clouds – a comparison
Model part | Implementation details |
|---|---|
Clustering | Filter and modified DBSCAN with parameters: \(\phantom {\dot {i}\!}\epsilon _{xyv_{r}} {=} 1.04, \epsilon _{v_{r}} {=} 1.03, \epsilon _{t} {=} 0.25, N_{\text {min,50}} {=} 3.87, v_{r,\text {min}} {=} 1.00, \alpha _{r} {=} 0.99\) |
Feature Extraction | 21 individually optimized feature vectors from a feature list of 98 handcrafted features (full list in [15]) |
Classification Ensemble | 15 OVO + 6 OVA classifiers with customized feature vectors |
LSTM Classifier | Single LSTM layer with 80 cells followed by a softmax layer, learning rate 10−3 |
Random Forest Classifier | 50 trees, Gini impurity, max split \(\sqrt {\text {feats.}}\), no restrictions on depth or split amount |
Class Proposal | Proposal equal to clusters, ensemble score defines class decision and confidence level |