Back to Portfolio
ai-ml

DeepTraject: Multi-Model Trajectory Prediction

A trajectory prediction framework that evaluates three different prediction approaches (Kalman Filter, LSTM, and Ensemble) across three distinct movement patterns: circular paths, figure-eight patterns, and lane-change maneuvers. This comprehensive evaluation demonstrates the strengths and limitations of each prediction method in different real-world scenarios.

Challenge

Create a unified prediction system that can accurately track and predict different types of movement patterns while comparing the effectiveness of classical, deep learning, and ensemble approaches across various trajectory types.

Solution

Developed a comparative framework that tests Kalman Filters, LSTM networks with attention, and an ensemble method across three specific movement patterns. Each approach is evaluated for its prediction accuracy, responsiveness, and adaptation to different trajectory types.

Implementation

  • Implemented three distinct test scenarios: circular paths, figure-eight patterns, and lane changes
  • Developed Kalman Filter tracking with real-time state estimation for each pattern
  • Created LSTM network with attention mechanism optimized for pattern recognition
  • Built ensemble system combining both approaches with dynamic weight adjustment
  • Implemented comprehensive visualization system showing real-time predictions
  • Created comparative analysis framework for all approaches across patterns

Simulation Analysis

Circular Pattern

Constant radius circular movement pattern testing continuous tracking ability

Circular Pattern Simulation
kalman
Accuracy: 96.5%Latency: 45ms
lstm
Accuracy: 95.8%Latency: 82ms
ensemble
Accuracy: 97.2%Latency: 88ms

Figure-Eight Pattern

Complex figure-eight pattern testing adaptation to changing curvature

Figure-Eight Pattern Simulation
kalman
Accuracy: 93.2%Latency: 48ms
lstm
Accuracy: 94.8%Latency: 85ms
ensemble
Accuracy: 95.4%Latency: 90ms

Lane Change Pattern

Sudden direction change testing rapid adaptation capabilities

Lane Change Pattern Simulation
kalman
Accuracy: 94.5%Latency: 42ms
lstm
Accuracy: 96.2%Latency: 80ms
ensemble
Accuracy: 96.5%Latency: 86ms

Technical Architecture

Test Pattern Comparison

Overview of the three test trajectories and prediction approaches

Test Pattern Comparison

Pattern-Specific Processing

Prediction process flow for each pattern type

Pattern-Specific Processing

Key Metrics

97.2%
Circular Pattern Accuracy
94.8%
Figure-8 Pattern Accuracy
96.5%
Lane Change Accuracy
<100ms
Update Speed

Project Links

View SourceDocumentation

Technologies

Python 3.8+PyTorchNumPyMatplotlibPlotlyAsyncio