I can suggest you a series network based on https://de.mathworks.com/help/vision/ug/point-cloud-classification-using-pointnet-deep-learning.html this documentation. But this is just a simplified example and should be adjusted to match your specific PointNet architecture.
% Define the layers for the PointNet model
inputLayer = imageInputLayer([3, 1, 1], 'Name', 'input', 'Normalization', 'none');
convLayer1 = convolution2dLayer(64, 1, 'Name', 'conv1', 'Padding', 'same');
convLayer2 = convolution2dLayer(64, 1, 'Name', 'conv2', 'Padding', 'same');
maxPoolLayer = maxPooling2dLayer([3, 1], 'Stride', [2, 1], 'Name', 'maxpool');
fullyConnectedLayer1 = fullyConnectedLayer(128, 'Name', 'fc1');
fullyConnectedLayer2 = fullyConnectedLayer(10, 'Name', 'fc2');
softmaxLayer = softmaxLayer('Name', 'softmax');
classificationLayer = classificationLayer('Name', 'classoutput');
% Assemble the layers into a Layer Graph
lgraph = layerGraph([inputLayer, convLayer1, reluLayer, convLayer2, reluLayer, maxPoolLayer, fullyConnectedLayer1, reluLayer, fullyConnectedLayer2, softmaxLayer, classificationLayer]);
% Convert the Layer Graph to a Series Network
net = assembleNetwork(lgraph);
