YUCHONG YE

The detection pipeline has three major steps:

​

1. At first, each input image frame is converted from RGB color space to HSV to better distinguish the car from background. Then HOG features, color histograms are extracted from each image and combined as a new feature vector. Since the magnitude of different features may vary a lot, the combine vector is then normalized. 

​

2. A linear SVM classifier is trained for on car and non-car image data.

​

3. For each video frame, I run a sliding window search to identify cars. There may be duplicate detections and false positives. So I built a heat-map to draw tighter bounding boxes and reject false positives. I also used a queue to store the car positions from previous 10 frames and calculated their average to smooth the detection results.