Home > Applications > Video > Video Analytics

Video Analytics Using CIM Accelerators

Vehicle and foot traffic recognition and tracking. Picture credit KAIST IPSL

Car detected in non-traffic area. Picture credit RaidNet

Surveillance video anomaly detection. Picture credit Butler University, Information Sciences and Systems

Overview
Applications
Source Code Examples

Overview

Video analytics, also sometimes referred to as Video Content Analysis, is the capability to automatically analyze video to detect and determine temporal events not based on a single image. That description is from Wikipedia, but it might be modified to include "patterns and objects from one or more images, and temporal events from multiple images or stream of images". Without fundamental pattern recognition and object detection, video content analysis cannot occur.

Applications

Video analytics, and its underlying foundation of object and pattern recognition, has many applications, including:

Security

Motion detection in a museum, store, bank, or other place secured for a period of time
Recognizing a person of interest in a surveillance video (known as face recognition, or facial recognition)

Robotics

Object tracking, object recognition
Location and mapping (sometimes referred to as SLAM, or Simultaneous Location and Mapping)

Military

Detecting recently disturbed earth or curbside debris along a road in order to identify a possible IED location
Recognizing unusual activity, such as vehicle or foot traffic out of the ordinary

Entertainment

Automatically recognizing a video's content, regardless of its labeling, to detect copyright infringement
Automatically recognizing and counting TV advertisements to ensure a radio or TV station has run ads they were paid to run by advertisers

Source Code Examples

Below are some video analytics source code examples, marked with CIM pragmas to accelerate compute-intensive sections of code. Note that CIM pragmas and API calls follow the OpenMP convention (highlighted in yellow below).


/* IIR Filter implementation from ffmpeg library, marked with CIM pragmas */

void ff_iir_filter(const struct FFIIRFilterCoeffs *c, struct FFIIRFilterState *s,
                   int size, const int16_t *src, int sstep, int16_t *dst, int dstep) {

int i;

   if (c->order == 4) {

      for (i=0; i<size; i+=4) {

         float in, res;

         FILTER(0, 1, 2, 3);
         FILTER(1, 2, 3, 0);
         FILTER(2, 3, 0, 1);
         FILTER(3, 0, 1, 2);
      }
      else {

         #pragma cim parallel for num_threads 4

         for (i=0; i<size; i++) {

         int j;
         float in, res;
         in = *src * c->gain;

         for (j=0; j<c->order; j++)
           in += c->cy[j] * s->x[j];

         res = s->x[0] + in + s->x[c->order >> 1] * c->cx[c->order >> 1];

         for (j=1; j<c->order>>1; j++)
           res += (s->x[j] + s->x[c->order - j]) * c->cx[j];

         for (j=0; j<c->order-1; j++)
           s->x[j] = s->x[j + 1];

         *dst = av_clip_int16(lrintf(res));

         s->x[c->order - 1] = in;

         src += sstep;
         dst += sstep;
      }
   }
}   

/* Discrete Cosine Transform (DCT) implementation from ffmpeg library, marked with CIM pragmas */
		
static void ff_dct_calc_III_c(DCTContext *ctx, FFTSample *data) {

int n = 1 << ctx->nbits;
int i;
float next = data[n - 1];
float inv_n = 1.0f / n;
		
   #pragma cim parallel for num_threads 6

   for (i=n-2; i>=2; i-=2) {

      float val1 = data[i];
      float val2 = data[i - 1] - data[i + 1];
      float c = COS(ctx, n, i);
      float s = SIN(ctx, n, i);

      data[i] = c * val1 + s * val2;
      data[i + 1] = s * val1 - c * val2;
   }
}